chore: bump version to 0.14.0 [skip ci]

Signed-off-by: 23Ro <m.n@23ro.de>

.github/dco.yml

@@ -0,0 +1,2 @@
+allowRemediationCommits:
+  individual: true

.github/workflows/ci-images-dryrun.yml

@@ -19,7 +19,7 @@ jobs:
             platforms: linux/amd64, linux/arm64
           - name: docling-project/docling-serve-cpu
             build_args: |
-              UV_SYNC_EXTRA_ARGS=--no-extra cu124
+              UV_SYNC_EXTRA_ARGS=--no-extra cu124 --no-extra flash-attn
             platforms: linux/amd64, linux/arm64
           - name: docling-project/docling-serve-cu124
             build_args: |

.github/workflows/dco-advisor.yml

@@ -0,0 +1,192 @@
+name: DCO Advisor Bot
+
+on:
+  pull_request_target:
+    types: [opened, reopened, synchronize]
+
+permissions:
+  pull-requests: write
+  issues: write
+
+jobs:
+  dco_advisor:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Handle DCO check result
+        uses: actions/github-script@v7
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const pr = context.payload.pull_request || context.payload.check_run?.pull_requests?.[0];
+            if (!pr) return;
+
+            const prNumber = pr.number;
+            const baseRef = pr.base.ref;
+            const headSha =
+              context.payload.check_run?.head_sha ||
+              pr.head?.sha;
+            const username = pr.user.login;
+
+            console.log("HEAD SHA:", headSha);
+
+            const sleep = ms => new Promise(resolve => setTimeout(resolve, ms));
+
+            // Poll until DCO check has a conclusion (max 6 attempts, 30s)
+            let dcoCheck = null;
+            for (let attempt = 0; attempt < 6; attempt++) {
+              const { data: checks } = await github.rest.checks.listForRef({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                ref: headSha
+              });
+
+              
+              console.log("All check runs:");
+                checks.check_runs.forEach(run => {
+                console.log(`- ${run.name} (${run.status}/${run.conclusion}) @ ${run.head_sha}`);
+              });
+
+              dcoCheck = checks.check_runs.find(run =>
+                run.name.toLowerCase().includes("dco") &&
+              !run.name.toLowerCase().includes("dco_advisor") &&
+                run.head_sha === headSha
+              );
+
+
+              if (dcoCheck?.conclusion) break;
+              console.log(`Waiting for DCO check... (${attempt + 1})`);
+              await sleep(5000); // wait 5 seconds
+            }
+
+            if (!dcoCheck || !dcoCheck.conclusion) {
+              console.log("DCO check did not complete in time.");
+              return;
+            }
+
+            const isFailure = ["failure", "action_required"].includes(dcoCheck.conclusion);
+            console.log(`DCO check conclusion for ${headSha}: ${dcoCheck.conclusion} (treated as ${isFailure ? "failure" : "success"})`);
+
+            // Parse DCO output for commit SHAs and author
+            let badCommits = [];
+            let authorName = "";
+            let authorEmail = "";
+            let moreInfo = `More info: [DCO check report](${dcoCheck?.html_url})`;
+
+            if (isFailure) {
+                const { data: commits } = await github.rest.pulls.listCommits({
+                    owner: context.repo.owner,
+                    repo: context.repo.repo,
+                    pull_number: prNumber,
+                });
+
+                for (const commit of commits) {
+                    const commitMessage = commit.commit.message;
+                    const signoffMatch = commitMessage.match(/^Signed-off-by:\s+.+<.+>$/m);
+                    if (!signoffMatch) {
+                        console.log(`Bad commit found ${commit.sha}`)
+                        badCommits.push({
+                        sha: commit.sha,
+                        authorName: commit.commit.author.name,
+                        authorEmail: commit.commit.author.email,
+                        });
+                    }
+                }            
+            }
+
+            // If multiple authors are present, you could adapt the message accordingly
+            // For now, we'll just use the first one
+            if (badCommits.length > 0) {
+            authorName = badCommits[0].authorName;
+            authorEmail = badCommits[0].authorEmail;
+            }
+
+            // Generate remediation commit message if needed
+            let remediationSnippet = "";
+            if (badCommits.length && authorEmail) {
+              remediationSnippet = `git commit --allow-empty -s -m "DCO Remediation Commit for ${authorName} <${authorEmail}>\n\n` +
+                badCommits.map(c => `I, ${c.authorName} <${c.authorEmail}>, hereby add my Signed-off-by to this commit: ${c.sha}`).join('\n') +
+                `"`;
+            } else {
+              remediationSnippet = "# Unable to auto-generate remediation message. Please check the DCO check details.";
+            }
+
+            // Build comment
+            const commentHeader = '<!-- dco-advice-bot -->';
+            let body = "";
+
+            if (isFailure) {
+              body = [
+                commentHeader,
+                '❌ **DCO Check Failed**',
+                '',
+                `Hi @${username}, your pull request has failed the Developer Certificate of Origin (DCO) check.`,
+                '',
+                'This repository supports **remediation commits**, so you can fix this without rewriting history — but you must follow the required message format.',
+                '',
+                '---',
+                '',
+                '### 🛠 Quick Fix: Add a remediation commit',
+                'Run this command:',
+                '',
+                '```bash',
+                remediationSnippet,
+                'git push',
+                '```',
+                '',
+                '---',
+                '',
+                '<details>',
+                '<summary>🔧 Advanced: Sign off each commit directly</summary>',
+                '',
+                '**For the latest commit:**',
+                '```bash',
+                'git commit --amend --signoff',
+                'git push --force-with-lease',
+                '```',
+                '',
+                '**For multiple commits:**',
+                '```bash',
+                `git rebase --signoff origin/${baseRef}`,
+                'git push --force-with-lease',
+                '```',
+                '',
+                '</details>',
+                '',
+                moreInfo
+              ].join('\n');
+            } else {
+              body = [
+                commentHeader,
+                '✅ **DCO Check Passed**',
+                '',
+                `Thanks @${username}, all your commits are properly signed off. 🎉`
+              ].join('\n');
+            }
+
+            // Get existing comments on the PR
+            const { data: comments } = await github.rest.issues.listComments({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: prNumber
+            });
+
+            // Look for a previous bot comment
+            const existingComment = comments.find(c =>
+              c.body.includes("<!-- dco-advice-bot -->")
+            );
+
+            if (existingComment) {
+              await github.rest.issues.updateComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                comment_id: existingComment.id,
+                body: body
+              });
+            } else {
+              await github.rest.issues.createComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: prNumber,
+                body: body
+              });
+            }

.github/workflows/images.yml

@@ -23,7 +23,7 @@ jobs:
             platforms: linux/amd64, linux/arm64
           - name: docling-project/docling-serve-cpu
             build_args: |
-              UV_SYNC_EXTRA_ARGS=--no-extra cu124
+              UV_SYNC_EXTRA_ARGS=--no-extra cu124 --no-extra flash-attn
             platforms: linux/amd64, linux/arm64
           - name: docling-project/docling-serve-cu124
             build_args: |

.github/workflows/job-build.yml

@@ -17,7 +17,7 @@ jobs:
           python-version: ${{ matrix.python-version }}
           enable-cache: true
       - name: Install dependencies
-        run: uv sync --all-extras --no-extra cu124
+        run: uv sync --all-extras --no-extra cu124 --no-extra flash-attn
       - name: Build package
         run: uv build
       - name: Check content of wheel

.github/workflows/job-checks.yml

@@ -25,7 +25,7 @@ jobs:
           key: pre-commit|${{ env.PY }}|${{ hashFiles('.pre-commit-config.yaml') }}
 
       - name: Install dependencies
-        run: uv sync --frozen --all-extras --no-extra cu124
+        run: uv sync --frozen --all-extras --no-extra cu124 --no-extra flash-attn
 
       - name: Run styling check
         run: pre-commit run --all-files

.gitignore

@@ -444,3 +444,5 @@ pip-selfcheck.json
 # Makefile
 .action-lint
 .markdown-lint
+
+cookies.txt

CHANGELOG.md

@@ -1,3 +1,87 @@
+## [v0.14.0](https://github.com/docling-project/docling-serve/releases/tag/v0.14.0) - 2025-06-17
+
+### Feature
+
+* Read supported file extensions from docling ([#214](https://github.com/docling-project/docling-serve/issues/214)) ([`524f6a8`](https://github.com/docling-project/docling-serve/commit/524f6a8997b86d2f869ca491ec8fb40585b42ca4))
+
+### Fix
+
+* Typo in Headline ([#220](https://github.com/docling-project/docling-serve/issues/220)) ([`d5455b7`](https://github.com/docling-project/docling-serve/commit/d5455b7f66de39ea1f8b8927b5968d2baa23ca88))
+
+## [v0.13.0](https://github.com/docling-project/docling-serve/releases/tag/v0.13.0) - 2025-06-04
+
+### Feature
+
+* Upgrade docling to 2.36 ([#212](https://github.com/docling-project/docling-serve/issues/212)) ([`ffea347`](https://github.com/docling-project/docling-serve/commit/ffea34732b24fdd438fabd6df02d3d9ce66b4534))
+
+## [v0.12.0](https://github.com/docling-project/docling-serve/releases/tag/v0.12.0) - 2025-06-03
+
+### Feature
+
+* Export annotations in markdown and html (Docling upgrade) ([#202](https://github.com/docling-project/docling-serve/issues/202)) ([`c4c41f1`](https://github.com/docling-project/docling-serve/commit/c4c41f16dff83c5d2a0b8a4c625b5de19b36b7c5))
+
+### Fix
+
+* Processing complex params in multipart-form ([#210](https://github.com/docling-project/docling-serve/issues/210)) ([`7066f35`](https://github.com/docling-project/docling-serve/commit/7066f3520a88c07df1c80a0cc6c4339eaac4d6a7))
+
+### Documentation
+
+* Add openshift replicasets examples ([#209](https://github.com/docling-project/docling-serve/issues/209)) ([`6a8190c`](https://github.com/docling-project/docling-serve/commit/6a8190c315792bd1e0e2b0af310656baaa5551e5))
+
+## [v0.11.0](https://github.com/docling-project/docling-serve/releases/tag/v0.11.0) - 2025-05-23
+
+### Feature
+
+* Page break placeholder in markdown exports options ([#194](https://github.com/docling-project/docling-serve/issues/194)) ([`32b8a80`](https://github.com/docling-project/docling-serve/commit/32b8a809f348bf9fbde657f93589a56935d3749d))
+* Clear results registry ([#192](https://github.com/docling-project/docling-serve/issues/192)) ([`de002df`](https://github.com/docling-project/docling-serve/commit/de002dfcdc111c942a08b156c84b7fa22b3fbaf3))
+* Upgrade to Docling 2.33.0 ([#198](https://github.com/docling-project/docling-serve/issues/198)) ([`abe5aa0`](https://github.com/docling-project/docling-serve/commit/abe5aa03f54d44ecf5c6d76e3258028997a53e68))
+* Api to trigger offloading the models ([#188](https://github.com/docling-project/docling-serve/issues/188)) ([`00be428`](https://github.com/docling-project/docling-serve/commit/00be4284904d55b78c75c5475578ef11c2ade94c))
+* Figure annotations @ docling components 0.0.7 ([#181](https://github.com/docling-project/docling-serve/issues/181)) ([`3ff1b2f`](https://github.com/docling-project/docling-serve/commit/3ff1b2f9834aca37472a895a0e3da47560457d77))
+
+### Fix
+
+* Usage of hashlib for FIPS ([#171](https://github.com/docling-project/docling-serve/issues/171)) ([`8406fb9`](https://github.com/docling-project/docling-serve/commit/8406fb9b59d83247b8379974cabed497703dfc4d))
+
+### Documentation
+
+* Example and instructions on how to load model weights to persistent volume ([#197](https://github.com/docling-project/docling-serve/issues/197)) ([`3f090b7`](https://github.com/docling-project/docling-serve/commit/3f090b7d15eaf696611d89bbbba5b98569610828))
+* Async api usage and fixes ([#195](https://github.com/docling-project/docling-serve/issues/195)) ([`21c1791`](https://github.com/docling-project/docling-serve/commit/21c1791e427f5b1946ed46c68dfda03c957dca8f))
+
+## [v0.10.1](https://github.com/docling-project/docling-serve/releases/tag/v0.10.1) - 2025-04-30
+
+### Fix
+
+* Avoid missing specialized keys in the options hash ([#166](https://github.com/docling-project/docling-serve/issues/166)) ([`36787bc`](https://github.com/docling-project/docling-serve/commit/36787bc0616356a6199da618d8646de51636b34e))
+* Allow users to set the area threshold for picture descriptions ([#165](https://github.com/docling-project/docling-serve/issues/165)) ([`509f488`](https://github.com/docling-project/docling-serve/commit/509f4889f8ed4c0f0ce25bec4126ef1f1199797c))
+* Expose max wait time in sync endpoints ([#164](https://github.com/docling-project/docling-serve/issues/164)) ([`919cf5c`](https://github.com/docling-project/docling-serve/commit/919cf5c0414f2f11eb8012f451fed7a8f582b7ad))
+* Add flash-attn for cuda images ([#161](https://github.com/docling-project/docling-serve/issues/161)) ([`35c2630`](https://github.com/docling-project/docling-serve/commit/35c2630c613cf229393fc67b6938152b063ff498))
+
+## [v0.10.0](https://github.com/docling-project/docling-serve/releases/tag/v0.10.0) - 2025-04-28
+
+### Feature
+
+* Add support for file upload and return as file in async endpoints ([#152](https://github.com/docling-project/docling-serve/issues/152)) ([`c65f3c6`](https://github.com/docling-project/docling-serve/commit/c65f3c654c76c6b64b6aada1f0a153d74789d629))
+
+### Documentation
+
+* Fix new default pdf_backend ([#158](https://github.com/docling-project/docling-serve/issues/158)) ([`829effe`](https://github.com/docling-project/docling-serve/commit/829effec1a1b80320ccaf2c501be8015169b6fa3))
+* Fixing small typo in docs ([#155](https://github.com/docling-project/docling-serve/issues/155)) ([`14bafb2`](https://github.com/docling-project/docling-serve/commit/14bafb26286b94f80b56846c50d6e9a6d99a9763))
+
+## [v0.9.0](https://github.com/docling-project/docling-serve/releases/tag/v0.9.0) - 2025-04-25
+
+### Feature
+
+* Expose picture description options ([#148](https://github.com/docling-project/docling-serve/issues/148)) ([`4c9571a`](https://github.com/docling-project/docling-serve/commit/4c9571a052d5ec0044e49225bc5615e13cdb0a56))
+* Add parameters for Kubeflow pipeline engine (WIP) ([#107](https://github.com/docling-project/docling-serve/issues/107)) ([`26bef5b`](https://github.com/docling-project/docling-serve/commit/26bef5bec060f0afd8d358816b68c3f2c0dd4bc2))
+
+### Fix
+
+* Produce image artifacts in referenced mode ([#151](https://github.com/docling-project/docling-serve/issues/151)) ([`71c5fae`](https://github.com/docling-project/docling-serve/commit/71c5fae505366459fd481d2ecdabc5ebed94d49c))
+
+### Documentation
+
+* Vlm and picture description options ([#149](https://github.com/docling-project/docling-serve/issues/149)) ([`91956cb`](https://github.com/docling-project/docling-serve/commit/91956cbf4e91cf82bb4d54ace397cdbbfaf594ba))
+
 ## [v0.8.0](https://github.com/docling-project/docling-serve/releases/tag/v0.8.0) - 2025-04-22
 
 ### Feature

Containerfile

@@ -46,7 +46,10 @@ RUN --mount=from=ghcr.io/astral-sh/uv:0.6.1,source=/uv,target=/bin/uv \
     --mount=type=cache,target=/opt/app-root/src/.cache/uv,uid=1001 \
     --mount=type=bind,source=uv.lock,target=uv.lock \
     --mount=type=bind,source=pyproject.toml,target=pyproject.toml \
-    umask 002 && uv sync --frozen --no-install-project --no-dev --all-extras ${UV_SYNC_EXTRA_ARGS}
+    umask 002 && \
+    UV_SYNC_ARGS="--frozen --no-install-project --no-dev --all-extras" && \
+    uv sync ${UV_SYNC_ARGS} ${UV_SYNC_EXTRA_ARGS} --no-extra flash-attn && \
+    FLASH_ATTENTION_SKIP_CUDA_BUILD=TRUE uv sync ${UV_SYNC_ARGS} ${UV_SYNC_EXTRA_ARGS} --no-build-isolation-package=flash-attn
 
 ARG MODELS_LIST="layout tableformer picture_classifier easyocr"
 

Makefile

@@ -35,7 +35,7 @@ docling-serve-image: Containerfile
 .PHONY: docling-serve-cpu-image
 docling-serve-cpu-image: Containerfile ## Build docling-serve "cpu only" container image
 	$(ECHO_PREFIX) printf "  %-12s Containerfile\n" "[docling-serve CPU]"
-	$(CMD_PREFIX) docker build --load --build-arg "UV_SYNC_EXTRA_ARGS=--no-extra cu124" -f Containerfile -t ghcr.io/docling-project/docling-serve-cpu:$(TAG) .
+	$(CMD_PREFIX) docker build --load --build-arg "UV_SYNC_EXTRA_ARGS=--no-extra cu124 --no-extra flash-attn" -f Containerfile -t ghcr.io/docling-project/docling-serve-cpu:$(TAG) .
 	$(CMD_PREFIX) docker tag ghcr.io/docling-project/docling-serve-cpu:$(TAG) ghcr.io/docling-project/docling-serve-cpu:$(BRANCH_TAG)
 	$(CMD_PREFIX) docker tag ghcr.io/docling-project/docling-serve-cpu:$(TAG) quay.io/docling-project/docling-serve-cpu:$(BRANCH_TAG)
 

README.md

@@ -70,7 +70,7 @@ An easy to use UI is available at the `/ui` endpoint.
 
 ## Documentation and advance usages
 
-Visit the [Docling Serve documentation](./docs/README.md) for learning how to [configure the webserver](./docs/configuration.md), use all the [runtime options](./docs/usage.md) of the API and [deployment examples](./docs/deployment.md).
+Visit the [Docling Serve documentation](./docs/README.md) for learning how to [configure the webserver](./docs/configuration.md), use all the [runtime options](./docs/usage.md) of the API and [deployment examples](./docs/deployment.md), pre-load model weights into a persistent volume [model weights on persistent volume](./docs/pre-loading-models.md)
 
 ## Get help and support
 

docling_serve/app.py

@@ -1,11 +1,11 @@
 import asyncio
 import importlib.metadata
 import logging
-import tempfile
+import shutil
+import time
 from contextlib import asynccontextmanager
 from io import BytesIO
-from pathlib import Path
-from typing import Annotated, Any, Optional, Union
+from typing import Annotated
 
 from fastapi import (
     BackgroundTasks,
@@ -28,31 +28,35 @@ from fastapi.staticfiles import StaticFiles
 
 from docling.datamodel.base_models import DocumentStream
 
+from docling_serve.datamodel.callback import (
+    ProgressCallbackRequest,
+    ProgressCallbackResponse,
+)
 from docling_serve.datamodel.convert import ConvertDocumentsOptions
 from docling_serve.datamodel.requests import (
     ConvertDocumentFileSourcesRequest,
+    ConvertDocumentHttpSourcesRequest,
     ConvertDocumentsRequest,
 )
 from docling_serve.datamodel.responses import (
+    ClearResponse,
     ConvertDocumentResponse,
     HealthCheckResponse,
     MessageKind,
     TaskStatusResponse,
     WebsocketMessage,
 )
-from docling_serve.docling_conversion import (
-    convert_documents,
-    get_converter,
-    get_pdf_pipeline_opts,
-)
-from docling_serve.engines import get_orchestrator
-from docling_serve.engines.async_local.orchestrator import (
-    AsyncLocalOrchestrator,
-    TaskNotFoundError,
+from docling_serve.datamodel.task import Task, TaskSource
+from docling_serve.docling_conversion import _get_converter_from_hash
+from docling_serve.engines.async_orchestrator import (
+    BaseAsyncOrchestrator,
+    ProgressInvalid,
 )
+from docling_serve.engines.async_orchestrator_factory import get_async_orchestrator
+from docling_serve.engines.base_orchestrator import TaskNotFoundError
 from docling_serve.helper_functions import FormDepends
-from docling_serve.response_preparation import process_results
 from docling_serve.settings import docling_serve_settings
+from docling_serve.storage import get_scratch
 
 
 # Set up custom logging as we'll be intermixes with FastAPI/Uvicorn's logging
@@ -90,11 +94,11 @@ _log = logging.getLogger(__name__)
 # Context manager to initialize and clean up the lifespan of the FastAPI app
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    # Converter with default options
-    pdf_format_option = get_pdf_pipeline_opts(ConvertDocumentsOptions())
-    get_converter(pdf_format_option)
+    orchestrator = get_async_orchestrator()
+    scratch_dir = get_scratch()
 
-    orchestrator = get_orchestrator()
+    # Warm up processing cache
+    await orchestrator.warm_up_caches()
 
     # Start the background queue processor
     queue_task = asyncio.create_task(orchestrator.process_queue())
@@ -108,6 +112,10 @@ async def lifespan(app: FastAPI):
     except asyncio.CancelledError:
         _log.info("Queue processor cancelled.")
 
+    # Remove scratch directory in case it was a tempfile
+    if docling_serve_settings.scratch_path is not None:
+        shutil.rmtree(scratch_dir, ignore_errors=True)
+
 
 ##################################
 # App creation and configuration #
@@ -157,7 +165,8 @@ def create_app():  # noqa: C901
 
             from docling_serve.gradio_ui import ui as gradio_ui
 
-            tmp_output_dir = Path(tempfile.mkdtemp())
+            tmp_output_dir = get_scratch() / "gradio"
+            tmp_output_dir.mkdir(exist_ok=True, parents=True)
             gradio_ui.gradio_output_dir = tmp_output_dir
             app = gr.mount_gradio_app(
                 app,
@@ -205,6 +214,55 @@ def create_app():  # noqa: C901
                 redoc_js_url="/static/redoc.standalone.js",
             )
 
+    ########################
+    # Async / Sync helpers #
+    ########################
+
+    async def _enque_source(
+        orchestrator: BaseAsyncOrchestrator, conversion_request: ConvertDocumentsRequest
+    ) -> Task:
+        sources: list[TaskSource] = []
+        if isinstance(conversion_request, ConvertDocumentFileSourcesRequest):
+            sources.extend(conversion_request.file_sources)
+        if isinstance(conversion_request, ConvertDocumentHttpSourcesRequest):
+            sources.extend(conversion_request.http_sources)
+
+        task = await orchestrator.enqueue(
+            sources=sources, options=conversion_request.options
+        )
+        return task
+
+    async def _enque_file(
+        orchestrator: BaseAsyncOrchestrator,
+        files: list[UploadFile],
+        options: ConvertDocumentsOptions,
+    ) -> Task:
+        _log.info(f"Received {len(files)} files for processing.")
+
+        # Load the uploaded files to Docling DocumentStream
+        file_sources: list[TaskSource] = []
+        for i, file in enumerate(files):
+            buf = BytesIO(file.file.read())
+            suffix = "" if len(file_sources) == 1 else f"_{i}"
+            name = file.filename if file.filename else f"file{suffix}.pdf"
+            file_sources.append(DocumentStream(name=name, stream=buf))
+
+        task = await orchestrator.enqueue(sources=file_sources, options=options)
+        return task
+
+    async def _wait_task_complete(
+        orchestrator: BaseAsyncOrchestrator, task_id: str
+    ) -> bool:
+        start_time = time.monotonic()
+        while True:
+            task = await orchestrator.task_status(task_id=task_id)
+            if task.is_completed():
+                return True
+            await asyncio.sleep(5)
+            elapsed_time = time.monotonic() - start_time
+            if elapsed_time > docling_serve_settings.max_sync_wait:
+                return False
+
     #############################
     # API Endpoints definitions #
     #############################
@@ -238,33 +296,34 @@ def create_app():  # noqa: C901
             }
         },
     )
-    def process_url(
-        background_tasks: BackgroundTasks, conversion_request: ConvertDocumentsRequest
+    async def process_url(
+        background_tasks: BackgroundTasks,
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
+        conversion_request: ConvertDocumentsRequest,
     ):
-        sources: list[Union[str, DocumentStream]] = []
-        headers: Optional[dict[str, Any]] = None
-        if isinstance(conversion_request, ConvertDocumentFileSourcesRequest):
-            for file_source in conversion_request.file_sources:
-                sources.append(file_source.to_document_stream())
-        else:
-            for http_source in conversion_request.http_sources:
-                sources.append(http_source.url)
-                if headers is None and http_source.headers:
-                    headers = http_source.headers
-
-        # Note: results are only an iterator->lazy evaluation
-        results = convert_documents(
-            sources=sources, options=conversion_request.options, headers=headers
+        task = await _enque_source(
+            orchestrator=orchestrator, conversion_request=conversion_request
+        )
+        success = await _wait_task_complete(
+            orchestrator=orchestrator, task_id=task.task_id
         )
 
-        # The real processing will happen here
-        response = process_results(
-            background_tasks=background_tasks,
-            conversion_options=conversion_request.options,
-            conv_results=results,
-        )
+        if not success:
+            # TODO: abort task!
+            return HTTPException(
+                status_code=504,
+                detail=f"Conversion is taking too long. The maximum wait time is configure as DOCLING_SERVE_MAX_SYNC_WAIT={docling_serve_settings.max_sync_wait}.",
+            )
 
-        return response
+        result = await orchestrator.task_result(
+            task_id=task.task_id, background_tasks=background_tasks
+        )
+        if result is None:
+            raise HTTPException(
+                status_code=404,
+                detail="Task result not found. Please wait for a completion status.",
+            )
+        return result
 
     # Convert a document from file(s)
     @app.post(
@@ -278,29 +337,35 @@ def create_app():  # noqa: C901
     )
     async def process_file(
         background_tasks: BackgroundTasks,
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
         files: list[UploadFile],
         options: Annotated[
             ConvertDocumentsOptions, FormDepends(ConvertDocumentsOptions)
         ],
     ):
-        _log.info(f"Received {len(files)} files for processing.")
-
-        # Load the uploaded files to Docling DocumentStream
-        file_sources = []
-        for file in files:
-            buf = BytesIO(file.file.read())
-            name = file.filename if file.filename else "file.pdf"
-            file_sources.append(DocumentStream(name=name, stream=buf))
-
-        results = convert_documents(sources=file_sources, options=options)
-
-        response = process_results(
-            background_tasks=background_tasks,
-            conversion_options=options,
-            conv_results=results,
+        task = await _enque_file(
+            orchestrator=orchestrator, files=files, options=options
+        )
+        success = await _wait_task_complete(
+            orchestrator=orchestrator, task_id=task.task_id
         )
 
-        return response
+        if not success:
+            # TODO: abort task!
+            return HTTPException(
+                status_code=504,
+                detail=f"Conversion is taking too long. The maximum wait time is configure as DOCLING_SERVE_MAX_SYNC_WAIT={docling_serve_settings.max_sync_wait}.",
+            )
+
+        result = await orchestrator.task_result(
+            task_id=task.task_id, background_tasks=background_tasks
+        )
+        if result is None:
+            raise HTTPException(
+                status_code=404,
+                detail="Task result not found. Please wait for a completion status.",
+            )
+        return result
 
     # Convert a document from URL(s) using the async api
     @app.post(
@@ -308,10 +373,12 @@ def create_app():  # noqa: C901
         response_model=TaskStatusResponse,
     )
     async def process_url_async(
-        orchestrator: Annotated[AsyncLocalOrchestrator, Depends(get_orchestrator)],
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
         conversion_request: ConvertDocumentsRequest,
     ):
-        task = await orchestrator.enqueue(request=conversion_request)
+        task = await _enque_source(
+            orchestrator=orchestrator, conversion_request=conversion_request
+        )
         task_queue_position = await orchestrator.get_queue_position(
             task_id=task.task_id
         )
@@ -319,6 +386,33 @@ def create_app():  # noqa: C901
             task_id=task.task_id,
             task_status=task.task_status,
             task_position=task_queue_position,
+            task_meta=task.processing_meta,
+        )
+
+    # Convert a document from file(s) using the async api
+    @app.post(
+        "/v1alpha/convert/file/async",
+        response_model=TaskStatusResponse,
+    )
+    async def process_file_async(
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
+        background_tasks: BackgroundTasks,
+        files: list[UploadFile],
+        options: Annotated[
+            ConvertDocumentsOptions, FormDepends(ConvertDocumentsOptions)
+        ],
+    ):
+        task = await _enque_file(
+            orchestrator=orchestrator, files=files, options=options
+        )
+        task_queue_position = await orchestrator.get_queue_position(
+            task_id=task.task_id
+        )
+        return TaskStatusResponse(
+            task_id=task.task_id,
+            task_status=task.task_status,
+            task_position=task_queue_position,
+            task_meta=task.processing_meta,
         )
 
     # Task status poll
@@ -327,7 +421,7 @@ def create_app():  # noqa: C901
         response_model=TaskStatusResponse,
     )
     async def task_status_poll(
-        orchestrator: Annotated[AsyncLocalOrchestrator, Depends(get_orchestrator)],
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
         task_id: str,
         wait: Annotated[
             float, Query(help="Number of seconds to wait for a completed status.")
@@ -342,6 +436,7 @@ def create_app():  # noqa: C901
             task_id=task.task_id,
             task_status=task.task_status,
             task_position=task_queue_position,
+            task_meta=task.processing_meta,
         )
 
     # Task status websocket
@@ -350,7 +445,7 @@ def create_app():  # noqa: C901
     )
     async def task_status_ws(
         websocket: WebSocket,
-        orchestrator: Annotated[AsyncLocalOrchestrator, Depends(get_orchestrator)],
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
         task_id: str,
     ):
         await websocket.accept()
@@ -375,6 +470,7 @@ def create_app():  # noqa: C901
                 task_id=task.task_id,
                 task_status=task.task_status,
                 task_position=task_queue_position,
+                task_meta=task.processing_meta,
             )
             await websocket.send_text(
                 WebsocketMessage(
@@ -389,6 +485,7 @@ def create_app():  # noqa: C901
                     task_id=task.task_id,
                     task_status=task.task_status,
                     task_position=task_queue_position,
+                    task_meta=task.processing_meta,
                 )
                 await websocket.send_text(
                     WebsocketMessage(
@@ -416,10 +513,13 @@ def create_app():  # noqa: C901
         },
     )
     async def task_result(
-        orchestrator: Annotated[AsyncLocalOrchestrator, Depends(get_orchestrator)],
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
+        background_tasks: BackgroundTasks,
         task_id: str,
     ):
-        result = await orchestrator.task_result(task_id=task_id)
+        result = await orchestrator.task_result(
+            task_id=task_id, background_tasks=background_tasks
+        )
         if result is None:
             raise HTTPException(
                 status_code=404,
@@ -427,4 +527,46 @@ def create_app():  # noqa: C901
             )
         return result
 
+    # Update task progress
+    @app.post(
+        "/v1alpha/callback/task/progress",
+        response_model=ProgressCallbackResponse,
+    )
+    async def callback_task_progress(
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
+        request: ProgressCallbackRequest,
+    ):
+        try:
+            await orchestrator.receive_task_progress(request=request)
+            return ProgressCallbackResponse(status="ack")
+        except TaskNotFoundError:
+            raise HTTPException(status_code=404, detail="Task not found.")
+        except ProgressInvalid as err:
+            raise HTTPException(
+                status_code=400, detail=f"Invalid progress payload: {err}"
+            )
+
+    #### Clear requests
+
+    # Offload models
+    @app.get(
+        "/v1alpha/clear/converters",
+        response_model=ClearResponse,
+    )
+    async def clear_converters():
+        _get_converter_from_hash.cache_clear()
+        return ClearResponse()
+
+    # Clean results
+    @app.get(
+        "/v1alpha/clear/results",
+        response_model=ClearResponse,
+    )
+    async def clear_results(
+        orchestrator: Annotated[BaseAsyncOrchestrator, Depends(get_async_orchestrator)],
+        older_then: float = 3600,
+    ):
+        await orchestrator.clear_results(older_than=older_then)
+        return ClearResponse()
+
     return app

docling_serve/datamodel/callback.py

@@ -0,0 +1,50 @@
+import enum
+from typing import Annotated, Literal
+
+from pydantic import BaseModel, Field
+
+
+class ProgressKind(str, enum.Enum):
+    SET_NUM_DOCS = "set_num_docs"
+    UPDATE_PROCESSED = "update_processed"
+
+
+class BaseProgress(BaseModel):
+    kind: ProgressKind
+
+
+class ProgressSetNumDocs(BaseProgress):
+    kind: Literal[ProgressKind.SET_NUM_DOCS] = ProgressKind.SET_NUM_DOCS
+
+    num_docs: int
+
+
+class SucceededDocsItem(BaseModel):
+    source: str
+
+
+class FailedDocsItem(BaseModel):
+    source: str
+    error: str
+
+
+class ProgressUpdateProcessed(BaseProgress):
+    kind: Literal[ProgressKind.UPDATE_PROCESSED] = ProgressKind.UPDATE_PROCESSED
+
+    num_processed: int
+    num_succeeded: int
+    num_failed: int
+
+    docs_succeeded: list[SucceededDocsItem]
+    docs_failed: list[FailedDocsItem]
+
+
+class ProgressCallbackRequest(BaseModel):
+    task_id: str
+    progress: Annotated[
+        ProgressSetNumDocs | ProgressUpdateProcessed, Field(discriminator="kind")
+    ]
+
+
+class ProgressCallbackResponse(BaseModel):
+    status: Literal["ack"] = "ack"

docling_serve/datamodel/convert.py

@@ -1,13 +1,15 @@
 # Define the input options for the API
-from typing import Annotated, Optional
+from typing import Annotated, Any, Optional
 
-from pydantic import BaseModel, Field
+from pydantic import AnyUrl, BaseModel, Field, model_validator
+from typing_extensions import Self
 
 from docling.datamodel.base_models import InputFormat, OutputFormat
 from docling.datamodel.pipeline_options import (
     EasyOcrOptions,
     PdfBackend,
     PdfPipeline,
+    PictureDescriptionBaseOptions,
     TableFormerMode,
     TableStructureOptions,
 )
@@ -26,6 +28,89 @@ ocr_factory = get_ocr_factory(
 ocr_engines_enum = ocr_factory.get_enum()
 
 
+class PictureDescriptionLocal(BaseModel):
+    repo_id: Annotated[
+        str,
+        Field(
+            description="Repository id from the Hugging Face Hub.",
+            examples=[
+                "HuggingFaceTB/SmolVLM-256M-Instruct",
+                "ibm-granite/granite-vision-3.2-2b",
+            ],
+        ),
+    ]
+    prompt: Annotated[
+        str,
+        Field(
+            description="Prompt used when calling the vision-language model.",
+            examples=[
+                "Describe this image in a few sentences.",
+                "This is a figure from a document. Provide a detailed description of it.",
+            ],
+        ),
+    ] = "Describe this image in a few sentences."
+    generation_config: Annotated[
+        dict[str, Any],
+        Field(
+            description="Config from https://huggingface.co/docs/transformers/en/main_classes/text_generation#transformers.GenerationConfig",
+            examples=[{"max_new_tokens": 200, "do_sample": False}],
+        ),
+    ] = {"max_new_tokens": 200, "do_sample": False}
+
+
+class PictureDescriptionApi(BaseModel):
+    url: Annotated[
+        AnyUrl,
+        Field(
+            description="Endpoint which accepts openai-api compatible requests.",
+            examples=[
+                AnyUrl(
+                    "http://localhost:8000/v1/chat/completions"
+                ),  # example of a local vllm api
+                AnyUrl(
+                    "http://localhost:11434/v1/chat/completions"
+                ),  # example of ollama
+            ],
+        ),
+    ]
+    headers: Annotated[
+        dict[str, str],
+        Field(
+            description="Headers used for calling the API endpoint. For example, it could include authentication headers."
+        ),
+    ] = {}
+    params: Annotated[
+        dict[str, Any],
+        Field(
+            description="Model parameters.",
+            examples=[
+                {  # on vllm
+                    "model": "HuggingFaceTB/SmolVLM-256M-Instruct",
+                    "max_completion_tokens": 200,
+                },
+                {  # on vllm
+                    "model": "ibm-granite/granite-vision-3.2-2b",
+                    "max_completion_tokens": 200,
+                },
+                {  # on ollama
+                    "model": "granite3.2-vision:2b"
+                },
+            ],
+        ),
+    ] = {}
+    timeout: Annotated[float, Field(description="Timeout for the API request.")] = 20
+    prompt: Annotated[
+        str,
+        Field(
+            description="Prompt used when calling the vision-language model.",
+            examples=[
+                "Describe this image in a few sentences.",
+                "This is a figures from a document. Provide a detailed description of it.",
+            ],
+        ),
+    ] = "Describe this image in a few sentences."
+
+
 class ConvertDocumentsOptions(BaseModel):
     from_formats: Annotated[
         list[InputFormat],
@@ -211,6 +296,14 @@ class ConvertDocumentsOptions(BaseModel):
         ),
     ] = 2.0
 
+    md_page_break_placeholder: Annotated[
+        str,
+        Field(
+            description="Add this placeholder betweek pages in the markdown output.",
+            examples=["<!-- page-break -->", ""],
+        ),
+    ] = ""
+
     do_code_enrichment: Annotated[
         bool,
         Field(
@@ -226,7 +319,7 @@ class ConvertDocumentsOptions(BaseModel):
         bool,
         Field(
             description=(
-                "If enabled, perform formula OCR, return Latex code. "
+                "If enabled, perform formula OCR, return LaTeX code. "
                 "Boolean. Optional, defaults to false."
             ),
             examples=[False],
@@ -254,3 +347,48 @@ class ConvertDocumentsOptions(BaseModel):
             examples=[False],
         ),
     ] = False
+
+    picture_description_area_threshold: Annotated[
+        float,
+        Field(
+            description="Minimum percentage of the area for a picture to be processed with the models.",
+            examples=[PictureDescriptionBaseOptions().picture_area_threshold],
+        ),
+    ] = PictureDescriptionBaseOptions().picture_area_threshold
+
+    picture_description_local: Annotated[
+        Optional[PictureDescriptionLocal],
+        Field(
+            description="Options for running a local vision-language model in the picture description. The parameters refer to a model hosted on Hugging Face. This parameter is mutually exclusive with picture_description_api.",
+            examples=[
+                PictureDescriptionLocal(repo_id="ibm-granite/granite-vision-3.2-2b"),
+                PictureDescriptionLocal(repo_id="HuggingFaceTB/SmolVLM-256M-Instruct"),
+            ],
+        ),
+    ] = None
+
+    picture_description_api: Annotated[
+        Optional[PictureDescriptionApi],
+        Field(
+            description="API details for using a vision-language model in the picture description. This parameter is mutually exclusive with picture_description_local.",
+            examples=[
+                PictureDescriptionApi(
+                    url="http://localhost:11434/v1/chat/completions",
+                    params={"model": "granite3.2-vision:2b"},
+                )
+            ],
+        ),
+    ] = None
+
+    @model_validator(mode="after")
+    def picture_description_exclusivity(self) -> Self:
+        # Validate picture description options
+        if (
+            self.picture_description_local is not None
+            and self.picture_description_api is not None
+        ):
+            raise ValueError(
+                "The parameters picture_description_local and picture_description_api are mutually exclusive, only one of them can be set."
+            )
+
+        return self

docling_serve/datamodel/engines.py

@@ -10,3 +10,4 @@ class TaskStatus(str, enum.Enum):
 
 class AsyncEngine(str, enum.Enum):
     LOCAL = "local"
+    KFP = "kfp"

docling_serve/datamodel/kfp.py

@@ -0,0 +1,7 @@
+from pydantic import AnyUrl, BaseModel
+
+
+class CallbackSpec(BaseModel):
+    url: AnyUrl
+    headers: dict[str, str] = {}
+    ca_cert: str = ""

docling_serve/datamodel/requests.py

@@ -2,7 +2,7 @@ import base64
 from io import BytesIO
 from typing import Annotated, Any, Union
 
-from pydantic import BaseModel, Field
+from pydantic import AnyHttpUrl, BaseModel, Field
 
 from docling.datamodel.base_models import DocumentStream
 
@@ -15,7 +15,7 @@ class DocumentsConvertBase(BaseModel):
 
 class HttpSource(BaseModel):
     url: Annotated[
-        str,
+        AnyHttpUrl,
         Field(
             description="HTTP url to process",
             examples=["https://arxiv.org/pdf/2206.01062"],

docling_serve/datamodel/responses.py

@@ -7,12 +7,18 @@ from docling.datamodel.document import ConversionStatus, ErrorItem
 from docling.utils.profiling import ProfilingItem
 from docling_core.types.doc import DoclingDocument
 
+from docling_serve.datamodel.task_meta import TaskProcessingMeta
+
 
 # Status
 class HealthCheckResponse(BaseModel):
     status: str = "ok"
 
 
+class ClearResponse(BaseModel):
+    status: str = "ok"
+
+
 class DocumentResponse(BaseModel):
     filename: str
     md_content: Optional[str] = None
@@ -38,6 +44,7 @@ class TaskStatusResponse(BaseModel):
     task_id: str
     task_status: str
     task_position: Optional[int] = None
+    task_meta: Optional[TaskProcessingMeta] = None
 
 
 class MessageKind(str, enum.Enum):

docling_serve/datamodel/task.py

@@ -1,17 +1,53 @@
-from typing import Optional
+import datetime
+from functools import partial
+from pathlib import Path
+from typing import Optional, Union
 
-from pydantic import BaseModel
+from fastapi.responses import FileResponse
+from pydantic import BaseModel, ConfigDict, Field
 
+from docling.datamodel.base_models import DocumentStream
+
+from docling_serve.datamodel.convert import ConvertDocumentsOptions
 from docling_serve.datamodel.engines import TaskStatus
-from docling_serve.datamodel.requests import ConvertDocumentsRequest
+from docling_serve.datamodel.requests import FileSource, HttpSource
 from docling_serve.datamodel.responses import ConvertDocumentResponse
+from docling_serve.datamodel.task_meta import TaskProcessingMeta
+
+TaskSource = Union[HttpSource, FileSource, DocumentStream]
 
 
 class Task(BaseModel):
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
     task_id: str
     task_status: TaskStatus = TaskStatus.PENDING
-    request: Optional[ConvertDocumentsRequest]
-    result: Optional[ConvertDocumentResponse] = None
+    sources: list[TaskSource] = []
+    options: Optional[ConvertDocumentsOptions]
+    result: Optional[Union[ConvertDocumentResponse, FileResponse]] = None
+    scratch_dir: Optional[Path] = None
+    processing_meta: Optional[TaskProcessingMeta] = None
+    created_at: datetime.datetime = Field(
+        default_factory=partial(datetime.datetime.now, datetime.timezone.utc)
+    )
+    started_at: Optional[datetime.datetime] = None
+    finished_at: Optional[datetime.datetime] = None
+    last_update_at: datetime.datetime = Field(
+        default_factory=partial(datetime.datetime.now, datetime.timezone.utc)
+    )
+
+    def set_status(self, status: TaskStatus):
+        now = datetime.datetime.now(datetime.timezone.utc)
+        if status == TaskStatus.STARTED and self.started_at is None:
+            self.started_at = now
+        if (
+            status in [TaskStatus.SUCCESS, TaskStatus.FAILURE]
+            and self.finished_at is None
+        ):
+            self.finished_at = now
+
+        self.last_update_at = now
+        self.task_status = status
 
     def is_completed(self) -> bool:
         if self.task_status in [TaskStatus.SUCCESS, TaskStatus.FAILURE]:

docling_serve/datamodel/task_meta.py

@@ -0,0 +1,8 @@
+from pydantic import BaseModel
+
+
+class TaskProcessingMeta(BaseModel):
+    num_docs: int
+    num_processed: int = 0
+    num_succeeded: int = 0
+    num_failed: int = 0

docling_serve/docling_conversion.py

@@ -21,6 +21,8 @@ from docling.datamodel.pipeline_options import (
     PdfBackend,
     PdfPipeline,
     PdfPipelineOptions,
+    PictureDescriptionApiOptions,
+    PictureDescriptionVlmOptions,
     TableFormerMode,
     VlmPipelineOptions,
     smoldocling_vlm_conversion_options,
@@ -40,15 +42,12 @@ _log = logging.getLogger(__name__)
 # Custom serializer for PdfFormatOption
 # (model_dump_json does not work with some classes)
 def _hash_pdf_format_option(pdf_format_option: PdfFormatOption) -> bytes:
-    data = pdf_format_option.model_dump()
+    data = pdf_format_option.model_dump(serialize_as_any=True)
 
     # pipeline_options are not fully serialized by model_dump, dedicated pass
     if pdf_format_option.pipeline_options:
-        data["pipeline_options"] = pdf_format_option.pipeline_options.model_dump()
-
-        # Replace `artifacts_path` with a string representation
-        data["pipeline_options"]["artifacts_path"] = repr(
-            data["pipeline_options"]["artifacts_path"]
+        data["pipeline_options"] = pdf_format_option.pipeline_options.model_dump(
+            serialize_as_any=True, mode="json"
         )
 
     # Replace `pipeline_cls` with a string representation
@@ -57,15 +56,11 @@ def _hash_pdf_format_option(pdf_format_option: PdfFormatOption) -> bytes:
     # Replace `backend` with a string representation
     data["backend"] = repr(data["backend"])
 
-    # Handle `device` in `accelerator_options`
-    if "accelerator_options" in data and "device" in data["accelerator_options"]:
-        data["accelerator_options"]["device"] = repr(
-            data["accelerator_options"]["device"]
-        )
-
     # Serialize the dictionary to JSON with sorted keys to have consistent hashes
     serialized_data = json.dumps(data, sort_keys=True)
-    options_hash = hashlib.sha1(serialized_data.encode()).digest()
+    options_hash = hashlib.sha1(
+        serialized_data.encode(), usedforsecurity=False
+    ).digest()
     return options_hash
 
 
@@ -116,6 +111,7 @@ def _parse_standard_pdf_opts(
 
     pipeline_options = PdfPipelineOptions(
         artifacts_path=artifacts_path,
+        enable_remote_services=docling_serve_settings.enable_remote_services,
         document_timeout=request.document_timeout,
         do_ocr=request.do_ocr,
         ocr_options=ocr_options,
@@ -129,9 +125,28 @@ def _parse_standard_pdf_opts(
 
     if request.image_export_mode != ImageRefMode.PLACEHOLDER:
         pipeline_options.generate_page_images = True
+        if request.image_export_mode == ImageRefMode.REFERENCED:
+            pipeline_options.generate_picture_images = True
         if request.images_scale:
             pipeline_options.images_scale = request.images_scale
 
+    if request.picture_description_local is not None:
+        pipeline_options.picture_description_options = (
+            PictureDescriptionVlmOptions.model_validate(
+                request.picture_description_local.model_dump()
+            )
+        )
+
+    if request.picture_description_api is not None:
+        pipeline_options.picture_description_options = (
+            PictureDescriptionApiOptions.model_validate(
+                request.picture_description_api.model_dump()
+            )
+        )
+    pipeline_options.picture_description_options.picture_area_threshold = (
+        request.picture_description_area_threshold
+    )
+
     return pipeline_options
 
 
@@ -179,7 +194,7 @@ def get_pdf_pipeline_opts(
     if docling_serve_settings.artifacts_path is not None:
         if str(docling_serve_settings.artifacts_path.absolute()) == "":
             _log.info(
-                "artifacts_path is an empty path, model weights will be dowloaded "
+                "artifacts_path is an empty path, model weights will be downloaded "
                 "at runtime."
             )
             artifacts_path = None

docling_serve/engines/__init__.py

@@ -1,8 +0,0 @@
-from functools import lru_cache
-
-from docling_serve.engines.async_local.orchestrator import AsyncLocalOrchestrator
-
-
-@lru_cache
-def get_orchestrator() -> AsyncLocalOrchestrator:
-    return AsyncLocalOrchestrator()

docling_serve/engines/async_kfp/kfp_pipeline.py

@@ -0,0 +1,137 @@
+# ruff: noqa: E402, UP006, UP035
+
+from typing import Any, Dict, List
+
+from kfp import dsl
+
+PYTHON_BASE_IMAGE = "python:3.12"
+
+
+@dsl.component(
+    base_image=PYTHON_BASE_IMAGE,
+    packages_to_install=[
+        "pydantic",
+        "docling-serve @ git+https://github.com/docling-project/docling-serve@feat-kfp-engine",
+    ],
+    pip_index_urls=["https://download.pytorch.org/whl/cpu", "https://pypi.org/simple"],
+)
+def generate_chunks(
+    run_name: str,
+    request: Dict[str, Any],
+    batch_size: int,
+    callbacks: List[Dict[str, Any]],
+) -> List[List[Dict[str, Any]]]:
+    from pydantic import TypeAdapter
+
+    from docling_serve.datamodel.callback import (
+        ProgressCallbackRequest,
+        ProgressSetNumDocs,
+    )
+    from docling_serve.datamodel.kfp import CallbackSpec
+    from docling_serve.engines.async_kfp.notify import notify_callbacks
+
+    CallbacksListType = TypeAdapter(list[CallbackSpec])
+
+    sources = request["http_sources"]
+    splits = [sources[i : i + batch_size] for i in range(0, len(sources), batch_size)]
+
+    total = sum(len(chunk) for chunk in splits)
+    payload = ProgressCallbackRequest(
+        task_id=run_name, progress=ProgressSetNumDocs(num_docs=total)
+    )
+    notify_callbacks(
+        payload=payload,
+        callbacks=CallbacksListType.validate_python(callbacks),
+    )
+
+    return splits
+
+
+@dsl.component(
+    base_image=PYTHON_BASE_IMAGE,
+    packages_to_install=[
+        "pydantic",
+        "docling-serve @ git+https://github.com/docling-project/docling-serve@feat-kfp-engine",
+    ],
+    pip_index_urls=["https://download.pytorch.org/whl/cpu", "https://pypi.org/simple"],
+)
+def convert_batch(
+    run_name: str,
+    data_splits: List[Dict[str, Any]],
+    request: Dict[str, Any],
+    callbacks: List[Dict[str, Any]],
+    output_path: dsl.OutputPath("Directory"),  # type: ignore
+):
+    from pathlib import Path
+
+    from pydantic import AnyUrl, TypeAdapter
+
+    from docling_serve.datamodel.callback import (
+        FailedDocsItem,
+        ProgressCallbackRequest,
+        ProgressUpdateProcessed,
+        SucceededDocsItem,
+    )
+    from docling_serve.datamodel.convert import ConvertDocumentsOptions
+    from docling_serve.datamodel.kfp import CallbackSpec
+    from docling_serve.datamodel.requests import HttpSource
+    from docling_serve.engines.async_kfp.notify import notify_callbacks
+
+    CallbacksListType = TypeAdapter(list[CallbackSpec])
+
+    convert_options = ConvertDocumentsOptions.model_validate(request["options"])
+    print(convert_options)
+
+    output_dir = Path(output_path)
+    output_dir.mkdir(exist_ok=True, parents=True)
+    docs_succeeded: list[SucceededDocsItem] = []
+    docs_failed: list[FailedDocsItem] = []
+    for source_dict in data_splits:
+        source = HttpSource.model_validate(source_dict)
+        filename = Path(str(AnyUrl(source.url).path)).name
+        output_filename = output_dir / filename
+        print(f"Writing {output_filename}")
+        with output_filename.open("w") as f:
+            f.write(source.model_dump_json())
+        docs_succeeded.append(SucceededDocsItem(source=source.url))
+
+    payload = ProgressCallbackRequest(
+        task_id=run_name,
+        progress=ProgressUpdateProcessed(
+            num_failed=len(docs_failed),
+            num_processed=len(docs_succeeded) + len(docs_failed),
+            num_succeeded=len(docs_succeeded),
+            docs_succeeded=docs_succeeded,
+            docs_failed=docs_failed,
+        ),
+    )
+
+    print(payload)
+    notify_callbacks(
+        payload=payload,
+        callbacks=CallbacksListType.validate_python(callbacks),
+    )
+
+
+@dsl.pipeline()
+def process(
+    batch_size: int,
+    request: Dict[str, Any],
+    callbacks: List[Dict[str, Any]] = [],
+    run_name: str = "",
+):
+    chunks_task = generate_chunks(
+        run_name=run_name,
+        request=request,
+        batch_size=batch_size,
+        callbacks=callbacks,
+    )
+    chunks_task.set_caching_options(False)
+
+    with dsl.ParallelFor(chunks_task.output, parallelism=4) as data_splits:
+        convert_batch(
+            run_name=run_name,
+            data_splits=data_splits,
+            request=request,
+            callbacks=callbacks,
+        )

docling_serve/engines/async_kfp/notify.py

@@ -0,0 +1,32 @@
+import ssl
+
+import certifi
+import httpx
+
+from docling_serve.datamodel.callback import ProgressCallbackRequest
+from docling_serve.datamodel.kfp import CallbackSpec
+
+
+def notify_callbacks(
+    payload: ProgressCallbackRequest,
+    callbacks: list[CallbackSpec],
+):
+    if len(callbacks) == 0:
+        return
+
+    for callback in callbacks:
+        # https://www.python-httpx.org/advanced/ssl/#configuring-client-instances
+        if callback.ca_cert:
+            ctx = ssl.create_default_context(cadata=callback.ca_cert)
+        else:
+            ctx = ssl.create_default_context(cafile=certifi.where())
+
+        try:
+            httpx.post(
+                str(callback.url),
+                headers=callback.headers,
+                json=payload.model_dump(mode="json"),
+                verify=ctx,
+            )
+        except httpx.HTTPError as err:
+            print(f"Error notifying callback {callback.url}: {err}")

docling_serve/engines/async_kfp/orchestrator.py

@@ -0,0 +1,235 @@
+import datetime
+import json
+import logging
+import uuid
+from pathlib import Path
+from typing import Optional
+
+from kfp_server_api.models import V2beta1RuntimeState
+from pydantic import BaseModel, TypeAdapter
+from pydantic_settings import SettingsConfigDict
+
+from docling_serve.datamodel.callback import (
+    ProgressCallbackRequest,
+    ProgressSetNumDocs,
+    ProgressUpdateProcessed,
+)
+from docling_serve.datamodel.convert import ConvertDocumentsOptions
+from docling_serve.datamodel.engines import TaskStatus
+from docling_serve.datamodel.kfp import CallbackSpec
+from docling_serve.datamodel.requests import HttpSource
+from docling_serve.datamodel.task import Task, TaskSource
+from docling_serve.datamodel.task_meta import TaskProcessingMeta
+from docling_serve.engines.async_kfp.kfp_pipeline import process
+from docling_serve.engines.async_orchestrator import (
+    BaseAsyncOrchestrator,
+    ProgressInvalid,
+)
+from docling_serve.settings import docling_serve_settings
+
+_log = logging.getLogger(__name__)
+
+
+class _RunItem(BaseModel):
+    model_config = SettingsConfigDict(arbitrary_types_allowed=True)
+
+    run_id: str
+    state: str
+    created_at: datetime.datetime
+    scheduled_at: datetime.datetime
+    finished_at: datetime.datetime
+
+
+class AsyncKfpOrchestrator(BaseAsyncOrchestrator):
+    def __init__(self):
+        super().__init__()
+        import kfp
+
+        kfp_endpoint = docling_serve_settings.eng_kfp_endpoint
+        if kfp_endpoint is None:
+            raise ValueError("KFP endpoint is required when using the KFP engine.")
+
+        kube_sa_token_path = Path("/run/secrets/kubernetes.io/serviceaccount/token")
+        kube_sa_ca_cert_path = Path(
+            "/run/secrets/kubernetes.io/serviceaccount/service-ca.crt"
+        )
+
+        ssl_ca_cert = docling_serve_settings.eng_kfp_ca_cert_path
+        token = docling_serve_settings.eng_kfp_token
+        if (
+            ssl_ca_cert is None
+            and ".svc" in kfp_endpoint.host
+            and kube_sa_ca_cert_path.exists()
+        ):
+            ssl_ca_cert = str(kube_sa_ca_cert_path)
+        if token is None and kube_sa_token_path.exists():
+            token = kube_sa_token_path.read_text()
+
+        self._client = kfp.Client(
+            host=str(kfp_endpoint),
+            existing_token=token,
+            ssl_ca_cert=ssl_ca_cert,
+            # verify_ssl=False,
+        )
+
+    async def enqueue(
+        self, sources: list[TaskSource], options: ConvertDocumentsOptions
+    ) -> Task:
+        callbacks = []
+        if docling_serve_settings.eng_kfp_self_callback_endpoint is not None:
+            headers = {}
+            if docling_serve_settings.eng_kfp_self_callback_token_path is not None:
+                token = (
+                    docling_serve_settings.eng_kfp_self_callback_token_path.read_text()
+                )
+                headers["Authorization"] = f"Bearer {token}"
+            ca_cert = ""
+            if docling_serve_settings.eng_kfp_self_callback_ca_cert_path is not None:
+                ca_cert = docling_serve_settings.eng_kfp_self_callback_ca_cert_path.read_text()
+            callbacks.append(
+                CallbackSpec(
+                    url=docling_serve_settings.eng_kfp_self_callback_endpoint,
+                    headers=headers,
+                    ca_cert=ca_cert,
+                )
+            )
+
+        CallbacksType = TypeAdapter(list[CallbackSpec])
+        SourcesListType = TypeAdapter(list[HttpSource])
+        http_sources = [s for s in sources if isinstance(s, HttpSource)]
+        # hack: since the current kfp backend is not resolving the job_id placeholder,
+        # we set the run_name and pass it as argument to the job itself.
+        run_name = f"docling-job-{uuid.uuid4()}"
+        kfp_run = self._client.create_run_from_pipeline_func(
+            process,
+            arguments={
+                "batch_size": 10,
+                "sources": SourcesListType.dump_python(http_sources, mode="json"),
+                "options": options.model_dump(mode="json"),
+                "callbacks": CallbacksType.dump_python(callbacks, mode="json"),
+                "run_name": run_name,
+            },
+            run_name=run_name,
+        )
+        task_id = kfp_run.run_id
+
+        task = Task(task_id=task_id, sources=sources, options=options)
+        await self.init_task_tracking(task)
+        return task
+
+    async def _update_task_from_run(self, task_id: str, wait: float = 0.0):
+        run_info = self._client.get_run(run_id=task_id)
+        task = await self.get_raw_task(task_id=task_id)
+        # RUNTIME_STATE_UNSPECIFIED = "RUNTIME_STATE_UNSPECIFIED"
+        # PENDING = "PENDING"
+        # RUNNING = "RUNNING"
+        # SUCCEEDED = "SUCCEEDED"
+        # SKIPPED = "SKIPPED"
+        # FAILED = "FAILED"
+        # CANCELING = "CANCELING"
+        # CANCELED = "CANCELED"
+        # PAUSED = "PAUSED"
+        if run_info.state == V2beta1RuntimeState.SUCCEEDED:
+            task.set_status(TaskStatus.SUCCESS)
+        elif run_info.state == V2beta1RuntimeState.PENDING:
+            task.set_status(TaskStatus.PENDING)
+        elif run_info.state == V2beta1RuntimeState.RUNNING:
+            task.set_status(TaskStatus.STARTED)
+        else:
+            task.set_status(TaskStatus.FAILURE)
+
+    async def task_status(self, task_id: str, wait: float = 0.0) -> Task:
+        await self._update_task_from_run(task_id=task_id, wait=wait)
+        return await self.get_raw_task(task_id=task_id)
+
+    async def _get_pending(self) -> list[_RunItem]:
+        runs: list[_RunItem] = []
+        next_page: Optional[str] = None
+        while True:
+            res = self._client.list_runs(
+                page_token=next_page,
+                page_size=20,
+                filter=json.dumps(
+                    {
+                        "predicates": [
+                            {
+                                "operation": "EQUALS",
+                                "key": "state",
+                                "stringValue": "PENDING",
+                            }
+                        ]
+                    }
+                ),
+            )
+            if res.runs is not None:
+                for run in res.runs:
+                    runs.append(
+                        _RunItem(
+                            run_id=run.run_id,
+                            state=run.state,
+                            created_at=run.created_at,
+                            scheduled_at=run.scheduled_at,
+                            finished_at=run.finished_at,
+                        )
+                    )
+            if res.next_page_token is None:
+                break
+            next_page = res.next_page_token
+        return runs
+
+    async def queue_size(self) -> int:
+        runs = await self._get_pending()
+        return len(runs)
+
+    async def get_queue_position(self, task_id: str) -> Optional[int]:
+        runs = await self._get_pending()
+        for pos, run in enumerate(runs, start=1):
+            if run.run_id == task_id:
+                return pos
+        return None
+
+    async def process_queue(self):
+        return
+
+    async def warm_up_caches(self):
+        return
+
+    async def _get_run_id(self, run_name: str) -> str:
+        res = self._client.list_runs(
+            filter=json.dumps(
+                {
+                    "predicates": [
+                        {
+                            "operation": "EQUALS",
+                            "key": "name",
+                            "stringValue": run_name,
+                        }
+                    ]
+                }
+            ),
+        )
+        if res.runs is not None and len(res.runs) > 0:
+            return res.runs[0].run_id
+        raise RuntimeError(f"Run with {run_name=} not found.")
+
+    async def receive_task_progress(self, request: ProgressCallbackRequest):
+        task_id = await self._get_run_id(run_name=request.task_id)
+        progress = request.progress
+        task = await self.get_raw_task(task_id=task_id)
+
+        if isinstance(progress, ProgressSetNumDocs):
+            task.processing_meta = TaskProcessingMeta(num_docs=progress.num_docs)
+            task.task_status = TaskStatus.STARTED
+
+        elif isinstance(progress, ProgressUpdateProcessed):
+            if task.processing_meta is None:
+                raise ProgressInvalid(
+                    "UpdateProcessed was called before setting the expected number of documents."
+                )
+            task.processing_meta.num_processed += progress.num_processed
+            task.processing_meta.num_succeeded += progress.num_succeeded
+            task.processing_meta.num_failed += progress.num_failed
+            task.task_status = TaskStatus.STARTED
+
+        # TODO: could be moved to BackgroundTask
+        await self.notify_task_subscribers(task_id=task_id)

docling_serve/engines/async_local/orchestrator.py

@@ -3,44 +3,30 @@ import logging
 import uuid
 from typing import Optional
 
-from fastapi import WebSocket
-
-from docling_serve.datamodel.engines import TaskStatus
-from docling_serve.datamodel.requests import ConvertDocumentsRequest
-from docling_serve.datamodel.responses import (
-    MessageKind,
-    TaskStatusResponse,
-    WebsocketMessage,
-)
-from docling_serve.datamodel.task import Task
+from docling_serve.datamodel.convert import ConvertDocumentsOptions
+from docling_serve.datamodel.task import Task, TaskSource
+from docling_serve.docling_conversion import get_converter, get_pdf_pipeline_opts
 from docling_serve.engines.async_local.worker import AsyncLocalWorker
-from docling_serve.engines.base_orchestrator import BaseOrchestrator
+from docling_serve.engines.async_orchestrator import BaseAsyncOrchestrator
 from docling_serve.settings import docling_serve_settings
 
 _log = logging.getLogger(__name__)
 
 
-class OrchestratorError(Exception):
-    pass
-
-
-class TaskNotFoundError(OrchestratorError):
-    pass
-
-
-class AsyncLocalOrchestrator(BaseOrchestrator):
+class AsyncLocalOrchestrator(BaseAsyncOrchestrator):
     def __init__(self):
+        super().__init__()
         self.task_queue = asyncio.Queue()
-        self.tasks: dict[str, Task] = {}
         self.queue_list: list[str] = []
-        self.task_subscribers: dict[str, set[WebSocket]] = {}
 
-    async def enqueue(self, request: ConvertDocumentsRequest) -> Task:
+    async def enqueue(
+        self, sources: list[TaskSource], options: ConvertDocumentsOptions
+    ) -> Task:
         task_id = str(uuid.uuid4())
-        task = Task(task_id=task_id, request=request)
-        self.tasks[task_id] = task
+        task = Task(task_id=task_id, sources=sources, options=options)
+        await self.init_task_tracking(task)
+
         self.queue_list.append(task_id)
-        self.task_subscribers[task_id] = set()
         await self.task_queue.put(task_id)
         return task
 
@@ -52,16 +38,6 @@ class AsyncLocalOrchestrator(BaseOrchestrator):
             self.queue_list.index(task_id) + 1 if task_id in self.queue_list else None
         )
 
-    async def task_status(self, task_id: str, wait: float = 0.0) -> Task:
-        if task_id not in self.tasks:
-            raise TaskNotFoundError()
-        return self.tasks[task_id]
-
-    async def task_result(self, task_id: str):
-        if task_id not in self.tasks:
-            raise TaskNotFoundError()
-        return self.tasks[task_id].result
-
     async def process_queue(self):
         # Create a pool of workers
         workers = []
@@ -75,28 +51,7 @@ class AsyncLocalOrchestrator(BaseOrchestrator):
         await asyncio.gather(*workers)
         _log.debug("All workers completed.")
 
-    async def notify_task_subscribers(self, task_id: str):
-        if task_id not in self.task_subscribers:
-            raise RuntimeError(f"Task {task_id} does not have a subscribers list.")
-
-        task = self.tasks[task_id]
-        task_queue_position = await self.get_queue_position(task_id)
-        msg = TaskStatusResponse(
-            task_id=task.task_id,
-            task_status=task.task_status,
-            task_position=task_queue_position,
-        )
-        for websocket in self.task_subscribers[task_id]:
-            await websocket.send_text(
-                WebsocketMessage(message=MessageKind.UPDATE, task=msg).model_dump_json()
-            )
-            if task.is_completed():
-                await websocket.close()
-
-    async def notify_queue_positions(self):
-        for task_id in self.task_subscribers.keys():
-            # notify only pending tasks
-            if self.tasks[task_id].task_status != TaskStatus.PENDING:
-                continue
-
-            await self.notify_task_subscribers(task_id)
+    async def warm_up_caches(self):
+        # Converter with default options
+        pdf_format_option = get_pdf_pipeline_opts(ConvertDocumentsOptions())
+        get_converter(pdf_format_option)

docling_serve/engines/async_local/worker.py

@@ -1,17 +1,18 @@
 import asyncio
 import logging
+import shutil
 import time
 from typing import TYPE_CHECKING, Any, Optional, Union
 
-from fastapi import BackgroundTasks
+from fastapi.responses import FileResponse
 
 from docling.datamodel.base_models import DocumentStream
 
 from docling_serve.datamodel.engines import TaskStatus
-from docling_serve.datamodel.requests import ConvertDocumentFileSourcesRequest
-from docling_serve.datamodel.responses import ConvertDocumentResponse
+from docling_serve.datamodel.requests import FileSource, HttpSource
 from docling_serve.docling_conversion import convert_documents
 from docling_serve.response_preparation import process_results
+from docling_serve.storage import get_scratch
 
 if TYPE_CHECKING:
     from docling_serve.engines.async_local.orchestrator import AsyncLocalOrchestrator
@@ -35,7 +36,7 @@ class AsyncLocalWorker:
             task = self.orchestrator.tasks[task_id]
 
             try:
-                task.task_status = TaskStatus.STARTED
+                task.set_status(TaskStatus.STARTED)
                 _log.info(f"Worker {self.worker_id} processing task {task_id}")
 
                 # Notify clients about task updates
@@ -44,61 +45,68 @@ class AsyncLocalWorker:
                 # Notify clients about queue updates
                 await self.orchestrator.notify_queue_positions()
 
-                # Get the current event loop
-                asyncio.get_event_loop()
-
                 # Define a callback function to send progress updates to the client.
                 # TODO: send partial updates, e.g. when a document in the batch is done
                 def run_conversion():
-                    sources: list[Union[str, DocumentStream]] = []
+                    convert_sources: list[Union[str, DocumentStream]] = []
                     headers: Optional[dict[str, Any]] = None
-                    if isinstance(task.request, ConvertDocumentFileSourcesRequest):
-                        for file_source in task.request.file_sources:
-                            sources.append(file_source.to_document_stream())
-                    else:
-                        for http_source in task.request.http_sources:
-                            sources.append(http_source.url)
-                            if headers is None and http_source.headers:
-                                headers = http_source.headers
+                    for source in task.sources:
+                        if isinstance(source, DocumentStream):
+                            convert_sources.append(source)
+                        elif isinstance(source, FileSource):
+                            convert_sources.append(source.to_document_stream())
+                        elif isinstance(source, HttpSource):
+                            convert_sources.append(str(source.url))
+                            if headers is None and source.headers:
+                                headers = source.headers
 
                     # Note: results are only an iterator->lazy evaluation
                     results = convert_documents(
-                        sources=sources,
-                        options=task.request.options,
+                        sources=convert_sources,
+                        options=task.options,
                         headers=headers,
                     )
 
                     # The real processing will happen here
+                    work_dir = get_scratch() / task_id
                     response = process_results(
-                        background_tasks=BackgroundTasks(),
-                        conversion_options=task.request.options,
+                        conversion_options=task.options,
                         conv_results=results,
+                        work_dir=work_dir,
                     )
 
+                    if work_dir.exists():
+                        task.scratch_dir = work_dir
+                        if not isinstance(response, FileResponse):
+                            _log.warning(
+                                f"Task {task_id=} produced content in {work_dir=} but the response is not a file."
+                            )
+                            shutil.rmtree(work_dir, ignore_errors=True)
+
                     return response
 
-                # Run the prediction in a thread to avoid blocking the event loop.
                 start_time = time.monotonic()
+
+                # Run the prediction in a thread to avoid blocking the event loop.
+                # Get the current event loop
+                # loop = asyncio.get_event_loop()
                 # future = asyncio.run_coroutine_threadsafe(
                 #     run_conversion(),
                 #     loop=loop
                 # )
                 # response = future.result()
 
+                # Run in a thread
                 response = await asyncio.to_thread(
                     run_conversion,
                 )
                 processing_time = time.monotonic() - start_time
 
-                if not isinstance(response, ConvertDocumentResponse):
-                    _log.error(
-                        f"Worker {self.worker_id} got un-processable "
-                        "result for {task_id}: {type(response)}"
-                    )
                 task.result = response
-                task.request = None
+                task.sources = []
+                task.options = None
 
-                task.task_status = TaskStatus.SUCCESS
+                task.set_status(TaskStatus.SUCCESS)
                 _log.info(
                     f"Worker {self.worker_id} completed job {task_id} "
                     f"in {processing_time:.2f} seconds"
@@ -108,7 +116,7 @@ class AsyncLocalWorker:
                 _log.error(
                     f"Worker {self.worker_id} failed to process job {task_id}: {e}"
                 )
-                task.task_status = TaskStatus.FAILURE
+                task.set_status(TaskStatus.FAILURE)
 
             finally:
                 await self.orchestrator.notify_task_subscribers(task_id)

docling_serve/engines/async_orchestrator.py

@@ -0,0 +1,127 @@
+import asyncio
+import datetime
+import logging
+import shutil
+from typing import Union
+
+from fastapi import BackgroundTasks, WebSocket
+from fastapi.responses import FileResponse
+
+from docling_serve.datamodel.callback import ProgressCallbackRequest
+from docling_serve.datamodel.engines import TaskStatus
+from docling_serve.datamodel.responses import (
+    ConvertDocumentResponse,
+    MessageKind,
+    TaskStatusResponse,
+    WebsocketMessage,
+)
+from docling_serve.datamodel.task import Task
+from docling_serve.engines.base_orchestrator import (
+    BaseOrchestrator,
+    OrchestratorError,
+    TaskNotFoundError,
+)
+from docling_serve.settings import docling_serve_settings
+
+_log = logging.getLogger(__name__)
+
+
+class ProgressInvalid(OrchestratorError):
+    pass
+
+
+class BaseAsyncOrchestrator(BaseOrchestrator):
+    def __init__(self):
+        self.tasks: dict[str, Task] = {}
+        self.task_subscribers: dict[str, set[WebSocket]] = {}
+
+    async def init_task_tracking(self, task: Task):
+        task_id = task.task_id
+        self.tasks[task.task_id] = task
+        self.task_subscribers[task_id] = set()
+
+    async def get_raw_task(self, task_id: str) -> Task:
+        if task_id not in self.tasks:
+            raise TaskNotFoundError()
+        return self.tasks[task_id]
+
+    async def task_status(self, task_id: str, wait: float = 0.0) -> Task:
+        return await self.get_raw_task(task_id=task_id)
+
+    async def task_result(
+        self, task_id: str, background_tasks: BackgroundTasks
+    ) -> Union[ConvertDocumentResponse, FileResponse, None]:
+        try:
+            task = await self.get_raw_task(task_id=task_id)
+            if task.is_completed() and docling_serve_settings.single_use_results:
+                if task.scratch_dir is not None:
+                    background_tasks.add_task(
+                        shutil.rmtree, task.scratch_dir, ignore_errors=True
+                    )
+
+                async def _remove_task_impl():
+                    await asyncio.sleep(docling_serve_settings.result_removal_delay)
+                    await self.delete_task(task_id=task.task_id)
+
+                async def _remove_task():
+                    asyncio.create_task(_remove_task_impl())  # noqa: RUF006
+
+                background_tasks.add_task(_remove_task)
+
+            return task.result
+        except TaskNotFoundError:
+            return None
+
+    async def delete_task(self, task_id: str):
+        _log.info(f"Deleting {task_id=}")
+        if task_id in self.task_subscribers:
+            for websocket in self.task_subscribers[task_id]:
+                await websocket.close()
+
+            del self.task_subscribers[task_id]
+
+        if task_id in self.tasks:
+            del self.tasks[task_id]
+
+    async def clear_results(self, older_than: float = 0.0):
+        cutoff_time = datetime.datetime.now(datetime.timezone.utc) - datetime.timedelta(
+            seconds=older_than
+        )
+
+        tasks_to_delete = [
+            task_id
+            for task_id, task in self.tasks.items()
+            if task.finished_at is not None and task.finished_at < cutoff_time
+        ]
+        for task_id in tasks_to_delete:
+            await self.delete_task(task_id=task_id)
+
+    async def notify_task_subscribers(self, task_id: str):
+        if task_id not in self.task_subscribers:
+            raise RuntimeError(f"Task {task_id} does not have a subscribers list.")
+
+        task = await self.get_raw_task(task_id=task_id)
+        task_queue_position = await self.get_queue_position(task_id)
+        msg = TaskStatusResponse(
+            task_id=task.task_id,
+            task_status=task.task_status,
+            task_position=task_queue_position,
+            task_meta=task.processing_meta,
+        )
+        for websocket in self.task_subscribers[task_id]:
+            await websocket.send_text(
+                WebsocketMessage(message=MessageKind.UPDATE, task=msg).model_dump_json()
+            )
+            if task.is_completed():
+                await websocket.close()
+
+    async def notify_queue_positions(self):
+        for task_id in self.task_subscribers.keys():
+            # notify only pending tasks
+            if self.tasks[task_id].task_status != TaskStatus.PENDING:
+                continue
+
+            await self.notify_task_subscribers(task_id)
+
+    async def receive_task_progress(self, request: ProgressCallbackRequest):
+        raise NotImplementedError()

docling_serve/engines/async_orchestrator_factory.py

@@ -0,0 +1,21 @@
+from functools import lru_cache
+
+from docling_serve.datamodel.engines import AsyncEngine
+from docling_serve.engines.async_orchestrator import BaseAsyncOrchestrator
+from docling_serve.settings import docling_serve_settings
+
+
+@lru_cache
+def get_async_orchestrator() -> BaseAsyncOrchestrator:
+    if docling_serve_settings.eng_kind == AsyncEngine.LOCAL:
+        from docling_serve.engines.async_local.orchestrator import (
+            AsyncLocalOrchestrator,
+        )
+
+        return AsyncLocalOrchestrator()
+    elif docling_serve_settings.eng_kind == AsyncEngine.KFP:
+        from docling_serve.engines.async_kfp.orchestrator import AsyncKfpOrchestrator
+
+        return AsyncKfpOrchestrator()
+
+    raise RuntimeError(f"Engine {docling_serve_settings.eng_kind} not recognized.")

docling_serve/engines/base_orchestrator.py

@@ -1,11 +1,27 @@
 from abc import ABC, abstractmethod
+from typing import Optional, Union
 
-from docling_serve.datamodel.task import Task
+from fastapi import BackgroundTasks
+from fastapi.responses import FileResponse
+
+from docling_serve.datamodel.convert import ConvertDocumentsOptions
+from docling_serve.datamodel.responses import ConvertDocumentResponse
+from docling_serve.datamodel.task import Task, TaskSource
+
+
+class OrchestratorError(Exception):
+    pass
+
+
+class TaskNotFoundError(OrchestratorError):
+    pass
 
 
 class BaseOrchestrator(ABC):
     @abstractmethod
-    async def enqueue(self, task) -> Task:
+    async def enqueue(
+        self, sources: list[TaskSource], options: ConvertDocumentsOptions
+    ) -> Task:
         pass
 
     @abstractmethod
@@ -13,9 +29,27 @@ class BaseOrchestrator(ABC):
         pass
 
     @abstractmethod
-    async def task_status(self, task_id: str) -> Task:
+    async def get_queue_position(self, task_id: str) -> Optional[int]:
         pass
 
     @abstractmethod
-    async def task_result(self, task_id: str):
+    async def task_status(self, task_id: str, wait: float = 0.0) -> Task:
+        pass
+
+    @abstractmethod
+    async def task_result(
+        self, task_id: str, background_tasks: BackgroundTasks
+    ) -> Union[ConvertDocumentResponse, FileResponse, None]:
+        pass
+
+    @abstractmethod
+    async def clear_results(self, older_than: float = 0.0):
+        pass
+
+    @abstractmethod
+    async def process_queue(self):
+        pass
+
+    @abstractmethod
+    async def warm_up_caches(self):
         pass

docling_serve/gradio_ui.py

@@ -1,16 +1,19 @@
 import base64
 import importlib
+import itertools
 import json
 import logging
 import ssl
 import tempfile
 import time
 from pathlib import Path
+from typing import Optional
 
 import certifi
 import gradio as gr
 import httpx
 
+from docling.datamodel.base_models import FormatToExtensions
 from docling.datamodel.pipeline_options import (
     PdfBackend,
     PdfPipeline,
@@ -28,7 +31,7 @@ logger = logging.getLogger(__name__)
 ############################
 
 logo_path = "https://raw.githubusercontent.com/docling-project/docling/refs/heads/main/docs/assets/logo.svg"
-js_components_url = "https://unpkg.com/@docling/docling-components@0.0.6"
+js_components_url = "https://unpkg.com/@docling/docling-components@0.0.7"
 if (
     docling_serve_settings.static_path is not None
     and docling_serve_settings.static_path.is_dir()
@@ -82,7 +85,7 @@ css = """
     height: 140px;
 }
 
-docling-img::part(pages) {
+docling-img {
     gap: 1rem;
 }
 
@@ -203,12 +206,16 @@ def clear_file_input():
     return None
 
 
-def auto_set_return_as_file(url_input, file_input, image_export_mode):
+def auto_set_return_as_file(
+    url_input_value: str,
+    file_input_value: Optional[list[str]],
+    image_export_mode_value: str,
+):
     # If more than one input source is provided, return as file
     if (
-        (len(url_input.split(",")) > 1)
-        or (file_input and len(file_input) > 1)
-        or (image_export_mode == "referenced")
+        (len(url_input_value.split(",")) > 1)
+        or (file_input_value and len(file_input_value) > 1)
+        or (image_export_mode_value == "referenced")
     ):
         return True
     else:
@@ -344,7 +351,7 @@ def file_to_base64(file):
 
 
 def process_file(
-    file,
+    files,
     to_formats,
     image_export_mode,
     pipeline,
@@ -361,10 +368,12 @@ def process_file(
     do_picture_classification,
     do_picture_description,
 ):
-    if not file or file == "":
+    if not files or len(files) == 0:
         logger.error("No files provided.")
         raise gr.Error("No files provided.", print_exception=False)
-    files_data = [{"base64_string": file_to_base64(file), "filename": file.name}]
+    files_data = [
+        {"base64_string": file_to_base64(file), "filename": file.name} for file in files
+    ]
 
     parameters = {
         "file_sources": files_data,
@@ -436,7 +445,7 @@ def response_to_output(response, return_as_file):
         )
         # Embed document JSON and trigger load at client via an image.
         json_rendered_content = f"""
-            <docling-img id="dclimg" pagenumbers tooltip="parsed"></docling-img>
+            <docling-img id="dclimg" pagenumbers><docling-tooltip></docling-tooltip></docling-img>
             <script id="dcljson" type="application/json" onload="document.getElementById('dclimg').src = JSON.parse(document.getElementById('dcljson').textContent);">{json_content}</script>
             <img src onerror="document.getElementById('dclimg').src = JSON.parse(document.getElementById('dcljson').textContent);" />
             """
@@ -538,21 +547,12 @@ with gr.Blocks(
                     elem_id="file_input_zone",
                     label="Upload File",
                     file_types=[
-                        ".pdf",
-                        ".docx",
-                        ".pptx",
-                        ".html",
-                        ".xlsx",
-                        ".json",
-                        ".asciidoc",
-                        ".txt",
-                        ".md",
-                        ".jpg",
-                        ".jpeg",
-                        ".png",
-                        ".gif",
+                        f".{v}"
+                        for v in itertools.chain.from_iterable(
+                            FormatToExtensions.values()
+                        )
                     ],
-                    file_count="single",
+                    file_count="multiple",
                     scale=4,
                 )
             with gr.Column(scale=1):
@@ -625,9 +625,7 @@ with gr.Blocks(
                 )
             with gr.Column(scale=1):
                 abort_on_error = gr.Checkbox(label="Abort on Error", value=False)
-                return_as_file = gr.Checkbox(
-                    label="Return as File", visible=False, value=False
-                )  # Disable until async handle output as file
+                return_as_file = gr.Checkbox(label="Return as File", value=False)
         with gr.Row():
             with gr.Column():
                 do_code_enrichment = gr.Checkbox(
@@ -677,23 +675,22 @@ with gr.Blocks(
     # UI Actions #
     ##############
 
-    # Disable until async handle output as file
     # Handle Return as File
-    # url_input.change(
-    #     auto_set_return_as_file,
-    #     inputs=[url_input, file_input, image_export_mode],
-    #     outputs=[return_as_file],
-    # )
-    # file_input.change(
-    #     auto_set_return_as_file,
-    #     inputs=[url_input, file_input, image_export_mode],
-    #     outputs=[return_as_file],
-    # )
-    # image_export_mode.change(
-    #     auto_set_return_as_file,
-    #     inputs=[url_input, file_input, image_export_mode],
-    #     outputs=[return_as_file],
-    # )
+    url_input.change(
+        auto_set_return_as_file,
+        inputs=[url_input, file_input, image_export_mode],
+        outputs=[return_as_file],
+    )
+    file_input.change(
+        auto_set_return_as_file,
+        inputs=[url_input, file_input, image_export_mode],
+        outputs=[return_as_file],
+    )
+    image_export_mode.change(
+        auto_set_return_as_file,
+        inputs=[url_input, file_input, image_export_mode],
+        outputs=[return_as_file],
+    )
 
     # URL processing
     url_process_btn.click(

docling_serve/helper_functions.py

@@ -1,9 +1,30 @@
 import inspect
+import json
 import re
-from typing import Union
+from typing import Union, get_args, get_origin
 
 from fastapi import Depends, Form
-from pydantic import BaseModel
+from pydantic import BaseModel, TypeAdapter
+
+
+def is_pydantic_model(type_):
+    try:
+        if inspect.isclass(type_) and issubclass(type_, BaseModel):
+            return True
+
+        origin = get_origin(type_)
+        if origin is Union:
+            args = get_args(type_)
+            return any(
+                inspect.isclass(arg) and issubclass(arg, BaseModel)
+                for arg in args
+                if arg is not type(None)
+            )
+
+    except Exception:
+        pass
+
+    return False
 
 
 # Adapted from
@@ -12,25 +33,62 @@ def FormDepends(cls: type[BaseModel]):
     new_parameters = []
 
     for field_name, model_field in cls.model_fields.items():
+        annotation = model_field.annotation
+        description = model_field.description
+        default = (
+            Form(..., description=description)
+            if model_field.is_required()
+            else Form(
+                model_field.default,
+                examples=model_field.examples,
+                description=description,
+            )
+        )
+
+        # Flatten nested Pydantic models by accepting them as JSON strings
+        if is_pydantic_model(annotation):
+            annotation = str
+            default = Form(
+                None
+                if model_field.default is None
+                else json.dumps(model_field.default.model_dump(mode="json")),
+                description=description,
+                examples=None
+                if not model_field.examples
+                else [
+                    json.dumps(ex.model_dump(mode="json"))
+                    for ex in model_field.examples
+                ],
+            )
+
         new_parameters.append(
             inspect.Parameter(
                 name=field_name,
                 kind=inspect.Parameter.POSITIONAL_ONLY,
-                default=(
-                    Form(...)
-                    if model_field.is_required()
-                    else Form(model_field.default)
-                ),
-                annotation=model_field.annotation,
+                default=default,
+                annotation=annotation,
             )
         )
 
     async def as_form_func(**data):
+        for field_name, model_field in cls.model_fields.items():
+            value = data.get(field_name)
+            annotation = model_field.annotation
+
+            # Parse nested models from JSON string
+            if value is not None and is_pydantic_model(annotation):
+                try:
+                    validator = TypeAdapter(annotation)
+                    data[field_name] = validator.validate_json(value)
+                except Exception as e:
+                    raise ValueError(f"Invalid JSON for field '{field_name}': {e}")
+
         return cls(**data)
 
     sig = inspect.signature(as_form_func)
     sig = sig.replace(parameters=new_parameters)
     as_form_func.__signature__ = sig  # type: ignore
+
     return Depends(as_form_func)
 
 

docling_serve/response_preparation.py

@@ -1,13 +1,12 @@
 import logging
 import os
 import shutil
-import tempfile
 import time
 from collections.abc import Iterable
 from pathlib import Path
 from typing import Union
 
-from fastapi import BackgroundTasks, HTTPException
+from fastapi import HTTPException
 from fastapi.responses import FileResponse
 
 from docling.datamodel.base_models import OutputFormat
@@ -28,6 +27,7 @@ def _export_document_as_content(
     export_txt: bool,
     export_doctags: bool,
     image_mode: ImageRefMode,
+    md_page_break_placeholder: str,
 ):
     document = DocumentResponse(filename=conv_res.input.file.name)
 
@@ -41,12 +41,16 @@ def _export_document_as_content(
             document.html_content = new_doc.export_to_html(image_mode=image_mode)
         if export_txt:
             document.text_content = new_doc.export_to_markdown(
-                strict_text=True, image_mode=image_mode
+                strict_text=True,
+                image_mode=image_mode,
             )
         if export_md:
-            document.md_content = new_doc.export_to_markdown(image_mode=image_mode)
+            document.md_content = new_doc.export_to_markdown(
+                image_mode=image_mode,
+                page_break_placeholder=md_page_break_placeholder or None,
+            )
         if export_doctags:
-            document.doctags_content = new_doc.export_to_document_tokens()
+            document.doctags_content = new_doc.export_to_doctags()
     elif conv_res.status == ConversionStatus.SKIPPED:
         raise HTTPException(status_code=400, detail=conv_res.errors)
     else:
@@ -64,6 +68,7 @@ def _export_documents_as_files(
     export_txt: bool,
     export_doctags: bool,
     image_export_mode: ImageRefMode,
+    md_page_break_placeholder: str,
 ):
     success_count = 0
     failure_count = 0
@@ -104,7 +109,9 @@ def _export_documents_as_files(
                 fname = output_dir / f"{doc_filename}.md"
                 _log.info(f"writing Markdown output to {fname}")
                 conv_res.document.save_as_markdown(
-                    filename=fname, image_mode=image_export_mode
+                    filename=fname,
+                    image_mode=image_export_mode,
+                    page_break_placeholder=md_page_break_placeholder or None,
                 )
 
             # Export Document Tags format:
@@ -124,9 +131,9 @@ def _export_documents_as_files(
 
 
 def process_results(
-    background_tasks: BackgroundTasks,
     conversion_options: ConvertDocumentsOptions,
     conv_results: Iterable[ConversionResult],
+    work_dir: Path,
 ) -> Union[ConvertDocumentResponse, FileResponse]:
     # Let's start by processing the documents
     try:
@@ -171,6 +178,7 @@ def process_results(
             export_txt=export_txt,
             export_doctags=export_doctags,
             image_mode=conversion_options.image_export_mode,
+            md_page_break_placeholder=conversion_options.md_page_break_placeholder,
         )
 
         response = ConvertDocumentResponse(
@@ -183,7 +191,6 @@ def process_results(
     # Multiple documents were processed, or we are forced returning as a file
     else:
         # Temporary directory to store the outputs
-        work_dir = Path(tempfile.mkdtemp(prefix="docling_"))
         output_dir = work_dir / "output"
         output_dir.mkdir(parents=True, exist_ok=True)
 
@@ -200,10 +207,10 @@ def process_results(
             export_txt=export_txt,
             export_doctags=export_doctags,
             image_export_mode=conversion_options.image_export_mode,
+            md_page_break_placeholder=conversion_options.md_page_break_placeholder,
         )
 
         files = os.listdir(output_dir)
-
         if len(files) == 0:
             raise HTTPException(status_code=500, detail="No documents were exported.")
 
@@ -216,7 +223,7 @@ def process_results(
 
         # Other cleanups after the response is sent
         # Output directory
-        background_tasks.add_task(shutil.rmtree, work_dir, ignore_errors=True)
+        # background_tasks.add_task(shutil.rmtree, work_dir, ignore_errors=True)
 
         response = FileResponse(
             file_path, filename=file_path.name, media_type="application/zip"

docling_serve/settings.py

@@ -2,7 +2,9 @@ import sys
 from pathlib import Path
 from typing import Optional, Union
 
+from pydantic import AnyUrl, model_validator
 from pydantic_settings import BaseSettings, SettingsConfigDict
+from typing_extensions import Self
 
 from docling_serve.datamodel.engines import AsyncEngine
 
@@ -36,19 +38,50 @@ class DoclingServeSettings(BaseSettings):
     api_host: str = "localhost"
     artifacts_path: Optional[Path] = None
     static_path: Optional[Path] = None
+    scratch_path: Optional[Path] = None
+    single_use_results: bool = True
+    result_removal_delay: float = 300  # 5 minutes
     options_cache_size: int = 2
+    enable_remote_services: bool = False
     allow_external_plugins: bool = False
 
     max_document_timeout: float = 3_600 * 24 * 7  # 7 days
     max_num_pages: int = sys.maxsize
     max_file_size: int = sys.maxsize
 
+    max_sync_wait: int = 120  # 2 minutes
+
     cors_origins: list[str] = ["*"]
     cors_methods: list[str] = ["*"]
     cors_headers: list[str] = ["*"]
 
     eng_kind: AsyncEngine = AsyncEngine.LOCAL
+    # Local engine
     eng_loc_num_workers: int = 2
+    # KFP engine
+    eng_kfp_endpoint: Optional[AnyUrl] = None
+    eng_kfp_token: Optional[str] = None
+    eng_kfp_ca_cert_path: Optional[str] = None
+    eng_kfp_self_callback_endpoint: Optional[str] = None
+    eng_kfp_self_callback_token_path: Optional[Path] = None
+    eng_kfp_self_callback_ca_cert_path: Optional[Path] = None
+
+    eng_kfp_experimental: bool = False
+
+    @model_validator(mode="after")
+    def engine_settings(self) -> Self:
+        # Validate KFP engine settings
+        if self.eng_kind == AsyncEngine.KFP:
+            if self.eng_kfp_endpoint is None:
+                raise ValueError("KFP endpoint is required when using the KFP engine.")
+
+        if self.eng_kind == AsyncEngine.KFP:
+            if not self.eng_kfp_experimental:
+                raise ValueError(
+                    "KFP is not yet working. To enable the development version, you must set DOCLING_SERVE_ENG_KFP_EXPERIMENTAL=true."
+                )
+
+        return self
 
 
 uvicorn_settings = UvicornSettings()

docling_serve/storage.py

@@ -0,0 +1,16 @@
+import tempfile
+from functools import lru_cache
+from pathlib import Path
+
+from docling_serve.settings import docling_serve_settings
+
+
+@lru_cache
+def get_scratch() -> Path:
+    scratch_dir = (
+        docling_serve_settings.scratch_path
+        if docling_serve_settings.scratch_path is not None
+        else Path(tempfile.mkdtemp(prefix="docling_"))
+    )
+    scratch_dir.mkdir(exist_ok=True, parents=True)
+    return scratch_dir

docs/README.md

@@ -1,4 +1,4 @@
-# Dolcing Serve documentation
+# Docling Serve documentation
 
 This documentation pages explore the webserver configurations, runtime options, deployment examples as well as development best practices.
 

docs/configuration.md

@@ -37,8 +37,44 @@ THe following table describes the options to configure the Docling Serve app.
 | -----------|-----|---------|-------------|
 | `--artifacts-path` | `DOCLING_SERVE_ARTIFACTS_PATH` | unset | If set to a valid directory, the model weights will be loaded from this path |
 |  | `DOCLING_SERVE_STATIC_PATH` | unset | If set to a valid directory, the static assets for the docs and ui will be loaded from this path |
+|  | `DOCLING_SERVE_SCRATCH_PATH` |  | If set, this directory will be used as scratch workspace, e.g. storing the results before they get requested. If unset, a temporary created is created for this purpose. |
 | `--enable-ui` | `DOCLING_SERVE_ENABLE_UI` | `false` | Enable the demonstrator UI. |
+|  | `DOCLING_SERVE_ENABLE_REMOTE_SERVICES` | `false` | Allow pipeline components making remote connections. For example, this is needed when using a vision-language model via APIs. |
+|  | `DOCLING_SERVE_ALLOW_EXTERNAL_PLUGINS` | `false` | Allow the selection of third-party plugins. |
+|  | `DOCLING_SERVE_SINGLE_USE_RESULTS` | `true` | If true, results can be accessed only once. If false, the results accumulate in the scratch directory. |
+|  | `DOCLING_SERVE_RESULT_REMOVAL_DELAY` | `300` | When `DOCLING_SERVE_SINGLE_USE_RESULTS` is active, this is the delay before results are removed from the task registry. |
+|  | `DOCLING_SERVE_MAX_DOCUMENT_TIMEOUT` | `604800` (7 days) | The maximum time for processing a document. |
+|  | `DOCLING_SERVE_MAX_NUM_PAGES` |  | The maximum number of pages for a document to be processed. |
+|  | `DOCLING_SERVE_MAX_FILE_SIZE` |  | The maximum file size for a document to be processed. |
+|  | `DOCLING_SERVE_MAX_SYNC_WAIT` | `120` | Max number of seconds a synchronous endpoint is waiting for the task completion. |
 |  | `DOCLING_SERVE_OPTIONS_CACHE_SIZE` | `2` | How many DocumentConveter objects (including their loaded models) to keep in the cache. |
 |  | `DOCLING_SERVE_CORS_ORIGINS` | `["*"]` | A list of origins that should be permitted to make cross-origin requests. |
 |  | `DOCLING_SERVE_CORS_METHODS` | `["*"]` | A list of HTTP methods that should be allowed for cross-origin requests. |
 |  | `DOCLING_SERVE_CORS_HEADERS` | `["*"]` | A list of HTTP request headers that should be supported for cross-origin requests. |
+|  | `DOCLING_SERVE_ENG_KIND` | `local` | The compute engine to use for the async tasks. Possible values are `local` and `kfp`. See below for more configurations of the engines. |
+
+### Compute engine
+
+Docling Serve can be deployed with several possible of compute engine.
+The selected compute engine will be running all the async jobs.
+
+#### Local engine
+
+The following table describes the options to configure the Docling Serve KFP engine.
+
+| ENV | Default | Description |
+|-----|---------|-------------|
+| `DOCLING_SERVE_ENG_LOC_NUM_WORKERS` | 2 | Number of workers/threads processing the incoming tasks. |
+
+#### KFP engine
+
+The following table describes the options to configure the Docling Serve KFP engine.
+
+| ENV | Default | Description |
+|-----|---------|-------------|
+| `DOCLING_SERVE_ENG_KFP_ENDPOINT` |  | Must be set to the Kubeflow Pipeline endpoint. When using the in-cluster deployment, make sure to use the cluster endpoint, e.g. `https://NAME.NAMESPACE.svc.cluster.local:8888`  |
+| `DOCLING_SERVE_ENG_KFP_TOKEN` |  | The authentication token for KFP. For in-cluster deployment, the app will load automatically the token of the ServiceAccount. |
+| `DOCLING_SERVE_ENG_KFP_CA_CERT_PATH` |  | Path to the CA certificates for the KFP endpoint. For in-cluster deployment, the app will load automatically the internal CA. |
+| `DOCLING_SERVE_ENG_KFP_SELF_CALLBACK_ENDPOINT` |  | If set, it enables internal callbacks providing status update of the KFP job. Usually something like `https://NAME.NAMESPACE.svc.cluster.local:5001/v1alpha/callback/task/progress`. |
+| `DOCLING_SERVE_ENG_KFP_SELF_CALLBACK_TOKEN_PATH` |  | The token used for authenticating the progress callback. For cluster-internal workloads, use `/run/secrets/kubernetes.io/serviceaccount/token`. |
+| `DOCLING_SERVE_ENG_KFP_SELF_CALLBACK_CA_CERT_PATH` |  | The CA certificate for the progress callback. For cluster-inetrnal workloads, use `/var/run/secrets/kubernetes.io/serviceaccount/service-ca.crt`. |

docs/deploy-examples/docling-model-cache-deployment.yaml

@@ -0,0 +1,47 @@
+kind: Deployment
+apiVersion: apps/v1
+metadata:
+  name: docling-serve
+  labels:
+    app: docling-serve
+    component: docling-serve-api
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: docling-serve
+      component: docling-serve-api
+  template:
+    metadata:
+      labels:
+        app: docling-serve
+        component: docling-serve-api
+    spec:
+      restartPolicy: Always
+      containers:
+        - name: api
+          resources:
+            limits:
+              cpu: 500m
+              memory: 2Gi
+            requests:
+              cpu: 250m
+              memory: 1Gi
+          env:
+            - name: DOCLING_SERVE_ENABLE_UI
+              value: 'true'
+            - name: DOCLING_SERVE_ARTIFACTS_PATH
+              value: '/modelcache'
+          ports:
+            - name: http
+              containerPort: 5001
+              protocol: TCP
+          imagePullPolicy: Always
+          image: 'ghcr.io/docling-project/docling-serve-cpu'
+          volumeMounts:
+            - name: docling-model-cache
+              mountPath: /modelcache
+      volumes:
+        - name: docling-model-cache
+          persistentVolumeClaim:
+            claimName: docling-model-cache-pvc

docs/deploy-examples/docling-model-cache-job.yaml

@@ -0,0 +1,33 @@
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: docling-model-cache-load
+spec:
+  selector: {}
+  template:
+    metadata:
+      name: docling-model-load
+    spec:
+      containers:
+        - name: loader
+          image: ghcr.io/docling-project/docling-serve-cpu:main
+          command:
+            - docling-tools
+            - models
+            - download
+            - '--output-dir=/modelcache'
+            - 'layout'
+            - 'tableformer'
+            - 'code_formula'
+            - 'picture_classifier'
+            - 'smolvlm'
+            - 'granite_vision'
+            - 'easyocr'
+          volumeMounts:
+            - name: docling-model-cache
+              mountPath: /modelcache
+      volumes:
+        - name: docling-model-cache
+          persistentVolumeClaim:
+            claimName: docling-model-cache-pvc
+      restartPolicy: Never

docs/deploy-examples/docling-model-cache-pvc.yaml

@@ -0,0 +1,11 @@
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: docling-model-cache-pvc
+spec:
+  accessModes:
+    - ReadWriteOnce
+  volumeMode: Filesystem
+  resources:
+    requests:
+      storage: 10Gi

docs/deploy-examples/docling-serve-replicas-w-sticky-sessions.yaml

@@ -0,0 +1,76 @@
+# This example deployment configures Docling Serve with a Route + Sticky sessions, a Service and cpu image
+---
+kind: Route
+apiVersion: route.openshift.io/v1
+metadata:
+  name: docling-serve
+  labels:
+    app: docling-serve
+    component: docling-serve-api
+  annotations:
+    haproxy.router.openshift.io/disable_cookies: "false" # this annotation enables the sticky sessions
+spec:
+  path: /
+  to:
+    kind: Service
+    name: docling-serve
+  port:
+    targetPort: http
+  tls:
+    termination: edge
+    insecureEdgeTerminationPolicy: Redirect
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: docling-serve
+  labels:
+    app: docling-serve
+    component: docling-serve-api
+spec:
+  ports:
+  - name: http
+    port: 5001
+    targetPort: http
+  selector:
+    app: docling-serve
+    component: docling-serve-api
+---
+kind: Deployment
+apiVersion: apps/v1
+metadata:
+  name: docling-serve
+  labels:
+    app: docling-serve
+    component: docling-serve-api
+spec:
+  replicas: 3
+  selector:
+    matchLabels:
+      app: docling-serve
+      component: docling-serve-api
+  template:
+    metadata:
+      labels:
+        app: docling-serve
+        component: docling-serve-api
+    spec:
+      restartPolicy: Always
+      containers:
+        - name: api
+          resources:
+            limits:
+              cpu: 500m
+              memory: 2Gi
+            requests:
+              cpu: 250m
+              memory: 1Gi
+          env:
+            - name: DOCLING_SERVE_ENABLE_UI
+              value: 'true'
+          ports:
+            - name: http
+              containerPort: 5001
+              protocol: TCP
+          imagePullPolicy: Always
+          image: 'ghcr.io/docling-project/docling-serve'

docs/deployment.md

@@ -192,3 +192,45 @@ curl -X 'POST' \
     "http_sources": [{"url": "https://arxiv.org/pdf/2501.17887"}]
   }'
 ```
+
+### ReplicaSets with `sticky sessions`
+
+Manifest example: [docling-serve-replicas-w-sticky-sessions.yaml](./deploy-examples/docling-serve-replicas-w-sticky-sessions.yaml)
+
+This deployment has the following features:
+
+- Deployment configuration with 3 replicas
+- Service configuration
+- Expose the service using a OpenShift `Route` and enables sticky sessions
+
+Install the app with:
+
+```sh
+oc apply -f docs/deploy-examples/docling-serve-replicas-w-sticky-sessions.yaml
+```
+
+For using the API:
+
+```sh
+# Retrieve the endpoint
+DOCLING_NAME=docling-serve
+DOCLING_ROUTE="https://$(oc get routes $DOCLING_NAME --template={{.spec.host}})"
+
+# Make a test query, store the cookie and taskid
+task_id=$(curl -s -X 'POST' \
+    "${DOCLING_ROUTE}/v1alpha/convert/source/async" \
+    -H "accept: application/json" \
+    -H "Content-Type: application/json" \
+    -d '{
+    "http_sources": [{"url": "https://arxiv.org/pdf/2501.17887"}]
+    }' \
+    -c cookies.txt | grep -oP '"task_id":"\K[^"]+')
+```
+
+```sh
+# Grab the taskid and cookie to check the task status
+curl -v -X 'GET' \
+  "${DOCLING_ROUTE}/v1alpha/status/poll/$task_id?wait=0" \
+  -H "accept: application/json" \
+  -b "cookies.txt"
+```

docs/pre-loading-models.md

@@ -0,0 +1,103 @@
+# Pre-loading models for docling
+
+This document provides examples for pre-loading docling models to a persistent volume and re-using it for docling-serve deployments.
+
+1. We need to create a persistent volume that will store models weights:
+
+    ```yaml
+    apiVersion: v1
+    kind: PersistentVolumeClaim
+    metadata:
+      name: docling-model-cache-pvc
+    spec:
+      accessModes:
+        - ReadWriteOnce
+      volumeMode: Filesystem
+      resources:
+        requests:
+          storage: 10Gi
+    ```
+
+    If you don't want to use default storage class, set your custom storage class with following:
+
+    ```yaml
+    spec:
+      ...
+      storageClassName: <Storage Class Name>
+    ```
+
+    Manifest example: [docling-model-cache-pvc.yaml](./deploy-examples/docling-model-cache-pvc.yaml)
+
+2. In order to load model weights, we can use docling-toolkit to download them, as this is a one time operation we can use kubernetes job for this:
+
+    ```yaml
+    apiVersion: batch/v1
+    kind: Job
+    metadata:
+      name: docling-model-cache-load
+    spec:
+      selector: {}
+      template:
+        metadata:
+          name: docling-model-load
+        spec:
+          containers:
+            - name: loader
+              image: ghcr.io/docling-project/docling-serve-cpu:main
+              command:
+                - docling-tools
+                - models
+                - download
+                - '--output-dir=/modelcache'
+                - 'layout'
+                - 'tableformer'
+                - 'code_formula'
+                - 'picture_classifier'
+                - 'smolvlm'
+                - 'granite_vision'
+                - 'easyocr'
+              volumeMounts:
+                - name: docling-model-cache
+                  mountPath: /modelcache
+          volumes:
+            - name: docling-model-cache
+              persistentVolumeClaim:
+                claimName: docling-model-cache-pvc
+          restartPolicy: Never
+    ```
+
+    The job will mount previously created persistent volume and execute command similar to how we would load models locally:
+    `docling-tools models download --output-dir <MOUNT-PATH> [LIST_OF_MODELS]`
+
+    In manifest, we specify desired models individually, or we can use `--all` parameter to download all models.
+
+    Manifest example: [docling-model-cache-job.yaml](./deploy-examples/docling-model-cache-job.yaml)
+
+3. Now we can mount volume in the docling-serve deployment and set env `DOCLING_SERVE_ARTIFACTS_PATH` to point to it.
+    Following additions to deploymeny should be made:
+
+    ```yaml
+    spec:
+      template:
+        spec:
+          containers:
+            - name: api
+              env:
+              ...
+                - name: DOCLING_SERVE_ARTIFACTS_PATH
+                  value: '/modelcache'
+              volumeMounts:
+                - name: docling-model-cache
+                  mountPath: /modelcache
+          ...
+          volumes:
+            - name: docling-model-cache
+              persistentVolumeClaim:
+                claimName: docling-model-cache-pvc
+    ```
+
+    Make sure that value of `DOCLING_SERVE_ARTIFACTS_PATH` is the same as where models were downloaded and where volume is mounted.
+
+    Now when docling-serve is executing tasks, the underlying docling installation will load model weights from mouted volume.
+
+    Manifest example: [docling-model-cache-deployment.yaml](./deploy-examples/docling-model-cache-deployment.yaml)

docs/usage.md

@@ -6,19 +6,29 @@ The API provides two endpoints: one for urls, one for files. This is necessary t
 
 On top of the source of file (see below), both endpoints support the same parameters, which are almost the same as the Docling CLI.
 
-- `from_format` (List[str]): Input format(s) to convert from. Allowed values: `docx`, `pptx`, `html`, `image`, `pdf`, `asciidoc`, `md`. Defaults to all formats.
+- `from_formats` (List[str]): Input format(s) to convert from. Allowed values: `docx`, `pptx`, `html`, `image`, `pdf`, `asciidoc`, `md`. Defaults to all formats.
 - `to_formats` (List[str]): Output format(s) to convert to. Allowed values: `md`, `json`, `html`, `text`, `doctags`. Defaults to `md`.
+- `pipeline` (str). The choice of which pipeline to use. Allowed values are `standard` and `vlm`. Defaults to `standard`.
+- `page_range` (tuple). If speficied, only convert a range of pages. The page number starts at 1.
 - `do_ocr` (bool): If enabled, the bitmap content will be processed using OCR. Defaults to `True`.
 - `image_export_mode`: Image export mode for the document (only in case of JSON, Markdown or HTML). Allowed values: embedded, placeholder, referenced. Optional, defaults to `embedded`.
 - `force_ocr` (bool): If enabled, replace any existing text with OCR-generated text over the full content. Defaults to `False`.
 - `ocr_engine` (str): OCR engine to use. Allowed values: `easyocr`, `tesseract_cli`, `tesseract`, `rapidocr`, `ocrmac`. Defaults to `easyocr`.
 - `ocr_lang` (List[str]): List of languages used by the OCR engine. Note that each OCR engine has different values for the language names. Defaults to empty.
-- `pdf_backend` (str): PDF backend to use. Allowed values: `pypdfium2`, `dlparse_v1`, `dlparse_v2`. Defaults to `dlparse_v2`.
+- `pdf_backend` (str): PDF backend to use. Allowed values: `pypdfium2`, `dlparse_v1`, `dlparse_v2`, `dlparse_v4`. Defaults to `dlparse_v4`.
 - `table_mode` (str): Table mode to use. Allowed values: `fast`, `accurate`. Defaults to `fast`.
 - `abort_on_error` (bool): If enabled, abort on error. Defaults to false.
 - `return_as_file` (boo): If enabled, return the output as a file. Defaults to false.
+- `md_page_break_placeholder` (str): Add this placeholder betweek pages in the markdown output.
 - `do_table_structure` (bool): If enabled, the table structure will be extracted. Defaults to true.
-- `include_images` (bool): If enabled, images will be extracted from the document. Defaults to true.
+- `do_code_enrichment` (bool): If enabled, perform OCR code enrichment. Defaults to false.
+- `do_formula_enrichment` (bool): If enabled, perform formula OCR, return LaTeX code. Defaults to false.
+- `do_picture_classification` (bool): If enabled, classify pictures in documents. Defaults to false.
+- `do_picture_description` (bool): If enabled, describe pictures in documents. Defaults to false.
+- `picture_description_area_threshold` (float): Minimum percentage of the area for a picture to be processed with the models. Defaults to 0.05.
+- `picture_description_local` (dict): Options for running a local vision-language model in the picture description. The parameters refer to a model hosted on Hugging Face. This parameter is mutually exclusive with picture_description_api.
+- `picture_description_api` (dict): API details for using a vision-language model in the picture description. This parameter is mutually exclusive with picture_description_local.
+- `include_images` (bool): If enabled, images will be extracted from the document. Defaults to false.
 - `images_scale` (float): Scale factor for images. Defaults to 2.0.
 
 ## Convert endpoints
@@ -236,7 +246,7 @@ files = {
     'files': ('2206.01062v1.pdf', open(file_path, 'rb'), 'application/pdf'),
 }
 
-response = await async_client.post(url, files=files, data={"parameters": json.dumps(parameters)})
+response = await async_client.post(url, files=files, data=parameters)
 assert response.status_code == 200, "Response should be 200 OK"
 
 data = response.json()
@@ -244,6 +254,70 @@ data = response.json()
 
 </details>
 
+### Picture description options
+
+When the picture description enrichment is activated, users may specify which model and which execution mode to use for this task. There are two choices for the execution mode: _local_ will run the vision-language model directly, _api_ will invoke an external API endpoint.
+
+The local option is specified with:
+
+```jsonc
+{
+  "picture_description_local": {
+    "repo_id": "",  // Repository id from the Hugging Face Hub.
+    "generation_config": {"max_new_tokens": 200, "do_sample": false},  // HF generation config.
+    "prompt": "Describe this image in a few sentences. ",  // Prompt used when calling the vision-language model.
+  }
+}
+```
+
+The possible values for `generation_config` are documented in the [Hugging Face text generation docs](https://huggingface.co/docs/transformers/en/main_classes/text_generation#transformers.GenerationConfig).
+
+The api option is specified with:
+
+```jsonc
+{
+  "picture_description_api": {
+    "url": "",  // Endpoint which accepts openai-api compatible requests.
+    "headers": {},  // Headers used for calling the API endpoint. For example, it could include authentication headers.
+    "params": {},  // Model parameters.
+    "timeout": 20,  // Timeout for the API request.
+    "prompt": "Describe this image in a few sentences. ",  // Prompt used when calling the vision-language model.
+  }
+}
+```
+
+Example URLs are:
+
+- `http://localhost:8000/v1/chat/completions` for the local vllm api, with example `params`:
+  - the `HuggingFaceTB/SmolVLM-256M-Instruct` model
+
+    ```json
+    {
+        "model": "HuggingFaceTB/SmolVLM-256M-Instruct",
+        "max_completion_tokens": 200,
+    }
+    ```
+  
+  - the `ibm-granite/granite-vision-3.2-2b` model
+
+    ```json
+    {
+        "model": "ibm-granite/granite-vision-3.2-2b",
+        "max_completion_tokens": 200,
+    }
+    ```
+
+- `http://localhost:11434/v1/chat/completions` for the local ollama api, with example `params`:
+  - the `granite3.2-vision:2b` model
+
+    ```json
+    {
+        "model": "granite3.2-vision:2b"
+    }
+    ```  
+
+Note that when using `picture_description_api`, the server must be launched with `DOCLING_SERVE_ENABLE_REMOTE_SERVICES=true`.
+
 ## Response format
 
 The response can be a JSON Document or a File.
@@ -276,4 +350,92 @@ The response can be a JSON Document or a File.
 
 ## Asynchronous API
 
-TBA
+Both `/v1alpha/convert/source` and `/v1alpha/convert/file` endpoints are available as asynchronous variants.
+The advantage of the asynchronous endpoints is the possible to interrupt the connection, check for the progress update and fetch the result.
+This approach is more resilient against network stabilities and allows the client application logic to easily interleave conversion with other tasks.
+
+Launch an asynchronous conversion with:
+
+- `POST /v1alpha/convert/source/async` when providing the input as sources.
+- `POST /v1alpha/convert/file/async` when providing the input as multipart-form files.
+
+The response format is a task detail:
+
+```jsonc
+{
+  "task_id": "<task_id>",  // the task_id which can be used for the next operations
+  "task_status": "pending|started|success|failure",  // the task status
+  "task_position": 1,  // the position in the queue
+  "task_meta": null,  // metadata e.g. how many documents are in the total job and how many have been converted
+}
+```
+
+### Polling status
+
+For checking the progress of the conversion task and wait for its completion, use the endpoint:
+
+- `GET /v1alpha/status/poll/{task_id}`
+
+<details>
+<summary>Example waiting loop:</summary>
+
+```python
+import time
+import httpx
+
+# ...
+# response from the async task submission
+task = response.json()
+
+while task["task_status"] not in ("success", "failure"):
+    response = httpx.get(f"{base_url}/status/poll/{task['task_id']}")
+    task = response.json()
+
+    time.sleep(5)
+```
+
+<details>
+
+### Subscribe with websockets
+
+Using websocket you can get the client application being notified about updates of the conversion task.
+To start the websocker connection, use the endpoint:
+
+- `/v1alpha/status/ws/{task_id}`
+
+Websocket messages are JSON object with the following structure:
+
+```jsonc
+{
+  "message": "connection|update|error",  // type of message being sent
+  "task": {},  // the same content of the task description
+  "error": "",  // description of the error
+}
+```
+
+<details>
+<summary>Example websocker usage:</summary>
+
+```python
+from websockets.sync.client import connect
+
+uri = f"ws://{base_url}/v1alpha/status/ws/{task['task_id']}"
+with connect(uri) as websocket:
+    for message in websocket:
+        try:
+            payload = json.loads(message)
+            if payload["message"] == "error":
+                break
+            if payload["message"] == "error" and payload["task"]["task_status"] in ("success", "failure"):
+                break
+        except:
+          break
+```
+
+</details>
+
+### Fetch results
+
+When the task is completed, the result can be fetched with the endpoint:
+
+- `GET /v1alpha/result/{task_id}`

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "docling-serve"
-version = "0.8.0"  # DO NOT EDIT, updated automatically
+version = "0.14.0"  # DO NOT EDIT, updated automatically
 description = "Running Docling as a service"
 license = {text = "MIT"}
 authors = [
@@ -31,9 +31,11 @@ classifiers = [
 requires-python = ">=3.10"
 dependencies = [
     "docling[vlm]~=2.28",
+    "docling-core>=2.32.0",
     "mlx-vlm~=0.1.12; sys_platform == 'darwin' and platform_machine == 'arm64'",
     "fastapi[standard]~=0.115",
     "httpx~=0.28",
+    "kfp[kubernetes]>=2.10.0",
     "pydantic~=2.10",
     "pydantic-settings~=2.4",
     "python-multipart>=0.0.14,<0.1.0",
@@ -62,9 +64,13 @@ cu124 = [
   "torch>=2.6.0",
   "torchvision>=0.21.0",
 ]
+flash-attn = [
+  "flash-attn~=2.7.0; sys_platform == 'linux' and platform_machine == 'x86_64'"
+]
 
 [dependency-groups]
 dev = [
+    "asgi-lifespan~=2.0",
     "mypy~=1.11",
     "pre-commit-uv~=4.1",
     "pytest~=8.3",
@@ -81,6 +87,13 @@ conflicts = [
     { extra = "cpu" },
     { extra = "cu124" },
   ],
+  [
+    { extra = "cpu" },
+    { extra = "flash-attn" },
+  ],]
+environments = ["sys_platform != 'darwin' or platform_machine != 'x86_64'"]
+override-dependencies = [
+  "urllib3~=2.0"
 ]
 
 [tool.uv.sources]
@@ -197,6 +210,8 @@ module = [
     "tesserocr.*",
     "rapidocr_onnxruntime.*",
     "requests.*",
+    "kfp.*",
+    "kfp_server_api.*",
     "mlx_vlm.*",
 ]
 ignore_missing_imports = true

tests/test_1-file-all-outputs.py

@@ -47,9 +47,7 @@ async def test_convert_file(async_client):
         "files": ("2206.01062v1.pdf", open(file_path, "rb"), "application/pdf"),
     }
 
-    response = await async_client.post(
-        url, files=files, data={"options": json.dumps(options)}
-    )
+    response = await async_client.post(url, files=files, data=options)
     assert response.status_code == 200, "Response should be 200 OK"
 
     data = response.json()

tests/test_1-file-async.py

@@ -0,0 +1,71 @@
+import json
+import time
+from pathlib import Path
+
+import httpx
+import pytest
+import pytest_asyncio
+
+
+@pytest_asyncio.fixture
+async def async_client():
+    async with httpx.AsyncClient(timeout=60.0) as client:
+        yield client
+
+
+@pytest.mark.asyncio
+async def test_convert_url(async_client):
+    """Test convert URL to all outputs"""
+
+    base_url = "http://localhost:5001/v1alpha"
+    payload = {
+        "to_formats": ["md", "json", "html"],
+        "image_export_mode": "placeholder",
+        "ocr": False,
+        "abort_on_error": False,
+        "return_as_file": False,
+    }
+
+    file_path = Path(__file__).parent / "2206.01062v1.pdf"
+    files = {
+        "files": (file_path.name, file_path.open("rb"), "application/pdf"),
+    }
+
+    for n in range(1):
+        response = await async_client.post(
+            f"{base_url}/convert/file/async", files=files, data=payload
+        )
+        assert response.status_code == 200, "Response should be 200 OK"
+
+    task = response.json()
+
+    print(json.dumps(task, indent=2))
+
+    while task["task_status"] not in ("success", "failure"):
+        response = await async_client.get(f"{base_url}/status/poll/{task['task_id']}")
+        assert response.status_code == 200, "Response should be 200 OK"
+        task = response.json()
+        print(f"{task['task_status']=}")
+        print(f"{task['task_position']=}")
+
+        time.sleep(2)
+
+    assert task["task_status"] == "success"
+    print(f"Task completed with status {task['task_status']=}")
+
+    result_resp = await async_client.get(f"{base_url}/result/{task['task_id']}")
+    assert result_resp.status_code == 200, "Response should be 200 OK"
+    result = result_resp.json()
+    print("Got result.")
+
+    assert "md_content" in result["document"]
+    assert result["document"]["md_content"] is not None
+    assert len(result["document"]["md_content"]) > 10
+
+    assert "html_content" in result["document"]
+    assert result["document"]["html_content"] is not None
+    assert len(result["document"]["html_content"]) > 10
+
+    assert "json_content" in result["document"]
+    assert result["document"]["json_content"] is not None
+    assert result["document"]["json_content"]["schema_name"] == "DoclingDocument"

tests/test_1-url-async-ws.py

@@ -28,6 +28,16 @@ async def test_convert_url(async_client: httpx.AsyncClient):
             "ocr": True,
             "abort_on_error": False,
             "return_as_file": False,
+            # "do_picture_description": True,
+            # "picture_description_api": {
+            #     "url": "http://localhost:11434/v1/chat/completions",
+            #     "params": {
+            #         "model": "granite3.2-vision:2b",
+            #     }
+            # },
+            # "picture_description_local": {
+            #     "repo_id": "HuggingFaceTB/SmolVLM-256M-Instruct",
+            # },
         },
         # "http_sources": [{"url": "https://arxiv.org/pdf/2501.17887"}],
         "file_sources": [{"base64_string": encoded_doc, "filename": doc_filename.name}],

tests/test_1-url-async.py

@@ -38,7 +38,7 @@ async def test_convert_url(async_client):
     }
     print(json.dumps(payload, indent=2))
 
-    for n in range(5):
+    for n in range(3):
         response = await async_client.post(
             f"{base_url}/convert/source/async", json=payload
         )

tests/test_2-files-all-outputs.py

@@ -1,4 +1,3 @@
-import json
 import os
 
 import httpx
@@ -48,9 +47,7 @@ async def test_convert_file(async_client):
         ("files", ("2408.09869v5.pdf", open(file_path, "rb"), "application/pdf")),
     ]
 
-    response = await async_client.post(
-        url, files=files, data={"options": json.dumps(options)}
-    )
+    response = await async_client.post(url, files=files, data=options)
     assert response.status_code == 200, "Response should be 200 OK"
 
     # Check for zip file attachment

tests/test_2-urls-async-all-outputs.py

@@ -0,0 +1,88 @@
+import json
+import time
+
+import httpx
+import pytest
+import pytest_asyncio
+from pytest_check import check
+
+
+@pytest_asyncio.fixture
+async def async_client():
+    async with httpx.AsyncClient(timeout=60.0) as client:
+        yield client
+
+
+@pytest.mark.asyncio
+async def test_convert_url(async_client):
+    """Test convert URL to all outputs"""
+    base_url = "http://localhost:5001/v1alpha"
+    payload = {
+        "options": {
+            "from_formats": [
+                "docx",
+                "pptx",
+                "html",
+                "image",
+                "pdf",
+                "asciidoc",
+                "md",
+                "xlsx",
+            ],
+            "to_formats": ["md", "json", "html", "text", "doctags"],
+            "image_export_mode": "placeholder",
+            "ocr": True,
+            "force_ocr": False,
+            "ocr_engine": "easyocr",
+            "ocr_lang": ["en"],
+            "pdf_backend": "dlparse_v2",
+            "table_mode": "fast",
+            "abort_on_error": False,
+            "return_as_file": False,
+        },
+        "http_sources": [
+            {"url": "https://arxiv.org/pdf/2206.01062"},
+            {"url": "https://arxiv.org/pdf/2408.09869"},
+        ],
+    }
+
+    response = await async_client.post(f"{base_url}/convert/source/async", json=payload)
+    assert response.status_code == 200, "Response should be 200 OK"
+
+    task = response.json()
+
+    print(json.dumps(task, indent=2))
+
+    while task["task_status"] not in ("success", "failure"):
+        response = await async_client.get(f"{base_url}/status/poll/{task['task_id']}")
+        assert response.status_code == 200, "Response should be 200 OK"
+        task = response.json()
+        print(f"{task['task_status']=}")
+        print(f"{task['task_position']=}")
+
+        time.sleep(2)
+
+    assert task["task_status"] == "success"
+
+    result_resp = await async_client.get(f"{base_url}/result/{task['task_id']}")
+    assert result_resp.status_code == 200, "Response should be 200 OK"
+
+    # Check for zip file attachment
+    content_disposition = result_resp.headers.get("content-disposition")
+
+    with check:
+        assert content_disposition is not None, (
+            "Content-Disposition header should be present"
+        )
+    with check:
+        assert "attachment" in content_disposition, "Response should be an attachment"
+    with check:
+        assert 'filename="converted_docs.zip"' in content_disposition, (
+            "Attachment filename should be 'converted_docs.zip'"
+        )
+
+    content_type = result_resp.headers.get("content-type")
+    with check:
+        assert content_type == "application/zip", (
+            "Content-Type should be 'application/zip'"
+        )

tests/test_fastapi_endpoints.py

@@ -0,0 +1,156 @@
+import asyncio
+import json
+import os
+
+import pytest
+import pytest_asyncio
+from asgi_lifespan import LifespanManager
+from httpx import ASGITransport, AsyncClient
+from pytest_check import check
+
+from docling_serve.app import create_app
+
+
+@pytest.fixture(scope="session")
+def event_loop():
+    return asyncio.get_event_loop()
+
+
+@pytest_asyncio.fixture(scope="session")
+async def app():
+    app = create_app()
+
+    async with LifespanManager(app) as manager:
+        print("Launching lifespan of app.")
+        yield manager.app
+
+
+@pytest_asyncio.fixture(scope="session")
+async def client(app):
+    async with AsyncClient(
+        transport=ASGITransport(app=app), base_url="http://app.io"
+    ) as client:
+        print("Client is ready")
+        yield client
+
+
+@pytest.mark.asyncio
+async def test_health(client: AsyncClient):
+    response = await client.get("/health")
+    assert response.status_code == 200
+    assert response.json() == {"status": "ok"}
+
+
+@pytest.mark.asyncio
+async def test_convert_file(client: AsyncClient):
+    """Test convert single file to all outputs"""
+
+    endpoint = "/v1alpha/convert/file"
+    options = {
+        "from_formats": [
+            "docx",
+            "pptx",
+            "html",
+            "image",
+            "pdf",
+            "asciidoc",
+            "md",
+            "xlsx",
+        ],
+        "to_formats": ["md", "json", "html", "text", "doctags"],
+        "image_export_mode": "placeholder",
+        "ocr": True,
+        "force_ocr": False,
+        "ocr_engine": "easyocr",
+        "ocr_lang": ["en"],
+        "pdf_backend": "dlparse_v2",
+        "table_mode": "fast",
+        "abort_on_error": False,
+        "return_as_file": False,
+    }
+
+    current_dir = os.path.dirname(__file__)
+    file_path = os.path.join(current_dir, "2206.01062v1.pdf")
+
+    files = {
+        "files": ("2206.01062v1.pdf", open(file_path, "rb"), "application/pdf"),
+    }
+
+    response = await client.post(endpoint, files=files, data=options)
+    assert response.status_code == 200, "Response should be 200 OK"
+
+    data = response.json()
+
+    # Response content checks
+    # Helper function to safely slice strings
+    def safe_slice(value, length=100):
+        if isinstance(value, str):
+            return value[:length]
+        return str(value)  # Convert non-string values to string for debug purposes
+
+    # Document check
+    check.is_in(
+        "document",
+        data,
+        msg=f"Response should contain 'document' key. Received keys: {list(data.keys())}",
+    )
+    # MD check
+    check.is_in(
+        "md_content",
+        data.get("document", {}),
+        msg=f"Response should contain 'md_content' key. Received keys: {list(data.get('document', {}).keys())}",
+    )
+    if data.get("document", {}).get("md_content") is not None:
+        check.is_in(
+            "## DocLayNet: ",
+            data["document"]["md_content"],
+            msg=f"Markdown document should contain 'DocLayNet: '. Received: {safe_slice(data['document']['md_content'])}",
+        )
+    # JSON check
+    check.is_in(
+        "json_content",
+        data.get("document", {}),
+        msg=f"Response should contain 'json_content' key. Received keys: {list(data.get('document', {}).keys())}",
+    )
+    if data.get("document", {}).get("json_content") is not None:
+        check.is_in(
+            '{"schema_name": "DoclingDocument"',
+            json.dumps(data["document"]["json_content"]),
+            msg=f'JSON document should contain \'{{\\n  "schema_name": "DoclingDocument\'". Received: {safe_slice(data["document"]["json_content"])}',
+        )
+    # HTML check
+    check.is_in(
+        "html_content",
+        data.get("document", {}),
+        msg=f"Response should contain 'html_content' key. Received keys: {list(data.get('document', {}).keys())}",
+    )
+    if data.get("document", {}).get("html_content") is not None:
+        check.is_in(
+            "<!DOCTYPE html>\n<html>\n<head>",
+            data["document"]["html_content"],
+            msg=f"HTML document should contain '<!DOCTYPE html>\n<html>\n<head>'. Received: {safe_slice(data['document']['html_content'])}",
+        )
+    # Text check
+    check.is_in(
+        "text_content",
+        data.get("document", {}),
+        msg=f"Response should contain 'text_content' key. Received keys: {list(data.get('document', {}).keys())}",
+    )
+    if data.get("document", {}).get("text_content") is not None:
+        check.is_in(
+            "DocLayNet: A Large Human-Annotated Dataset",
+            data["document"]["text_content"],
+            msg=f"Text document should contain 'DocLayNet: A Large Human-Annotated Dataset'. Received: {safe_slice(data['document']['text_content'])}",
+        )
+    # DocTags check
+    check.is_in(
+        "doctags_content",
+        data.get("document", {}),
+        msg=f"Response should contain 'doctags_content' key. Received keys: {list(data.get('document', {}).keys())}",
+    )
+    if data.get("document", {}).get("doctags_content") is not None:
+        check.is_in(
+            "<doctag><page_header>",
+            data["document"]["doctags_content"],
+            msg=f"DocTags document should contain '<doctag><page_header>'. Received: {safe_slice(data['document']['doctags_content'])}",
+        )

tests/test_file_opts.py

@@ -0,0 +1,77 @@
+import asyncio
+import json
+import os
+
+import pytest
+import pytest_asyncio
+from asgi_lifespan import LifespanManager
+from httpx import ASGITransport, AsyncClient
+
+from docling_core.types import DoclingDocument
+from docling_core.types.doc.document import PictureDescriptionData
+
+from docling_serve.app import create_app
+
+
+@pytest.fixture(scope="session")
+def event_loop():
+    return asyncio.get_event_loop()
+
+
+@pytest_asyncio.fixture(scope="session")
+async def app():
+    app = create_app()
+
+    async with LifespanManager(app) as manager:
+        print("Launching lifespan of app.")
+        yield manager.app
+
+
+@pytest_asyncio.fixture(scope="session")
+async def client(app):
+    async with AsyncClient(
+        transport=ASGITransport(app=app), base_url="http://app.io"
+    ) as client:
+        print("Client is ready")
+        yield client
+
+
+@pytest.mark.asyncio
+async def test_convert_file(client: AsyncClient):
+    """Test convert single file to all outputs"""
+
+    endpoint = "/v1alpha/convert/file"
+    options = {
+        "to_formats": ["md", "json"],
+        "image_export_mode": "placeholder",
+        "ocr": False,
+        "do_picture_description": True,
+        "picture_description_api": json.dumps(
+            {
+                "url": "http://localhost:11434/v1/chat/completions",  # ollama
+                "params": {"model": "granite3.2-vision:2b"},
+                "timeout": 60,
+                "prompt": "Describe this image in a few sentences. ",
+            }
+        ),
+    }
+
+    current_dir = os.path.dirname(__file__)
+    file_path = os.path.join(current_dir, "2206.01062v1.pdf")
+
+    files = {
+        "files": ("2206.01062v1.pdf", open(file_path, "rb"), "application/pdf"),
+    }
+
+    response = await client.post(endpoint, files=files, data=options)
+    assert response.status_code == 200, "Response should be 200 OK"
+
+    data = response.json()
+
+    doc = DoclingDocument.model_validate(data["document"]["json_content"])
+
+    for pic in doc.pictures:
+        for ann in pic.annotations:
+            if isinstance(ann, PictureDescriptionData):
+                print(f"{pic.self_ref}")
+                print(ann.text)

tests/test_options_serialization.py

@@ -0,0 +1,54 @@
+from docling_serve.datamodel.convert import (
+    ConvertDocumentsOptions,
+    PictureDescriptionApi,
+)
+from docling_serve.docling_conversion import (
+    _hash_pdf_format_option,
+    get_pdf_pipeline_opts,
+)
+
+
+def test_options_cache_key():
+    hashes = set()
+
+    opts = ConvertDocumentsOptions()
+    pipeline_opts = get_pdf_pipeline_opts(opts)
+    hash = _hash_pdf_format_option(pipeline_opts)
+    assert hash not in hashes
+    hashes.add(hash)
+
+    opts.do_picture_description = True
+    pipeline_opts = get_pdf_pipeline_opts(opts)
+    hash = _hash_pdf_format_option(pipeline_opts)
+    # pprint(pipeline_opts.pipeline_options.model_dump(serialize_as_any=True))
+    assert hash not in hashes
+    hashes.add(hash)
+
+    opts.picture_description_api = PictureDescriptionApi(
+        url="http://localhost",
+        params={"model": "mymodel"},
+        prompt="Hello 1",
+    )
+    pipeline_opts = get_pdf_pipeline_opts(opts)
+    hash = _hash_pdf_format_option(pipeline_opts)
+    # pprint(pipeline_opts.pipeline_options.model_dump(serialize_as_any=True))
+    assert hash not in hashes
+    hashes.add(hash)
+
+    opts.picture_description_api = PictureDescriptionApi(
+        url="http://localhost",
+        params={"model": "your-model"},
+        prompt="Hello 1",
+    )
+    pipeline_opts = get_pdf_pipeline_opts(opts)
+    hash = _hash_pdf_format_option(pipeline_opts)
+    # pprint(pipeline_opts.pipeline_options.model_dump(serialize_as_any=True))
+    assert hash not in hashes
+    hashes.add(hash)
+
+    opts.picture_description_api.prompt = "World"
+    pipeline_opts = get_pdf_pipeline_opts(opts)
+    hash = _hash_pdf_format_option(pipeline_opts)
+    # pprint(pipeline_opts.pipeline_options.model_dump(serialize_as_any=True))
+    assert hash not in hashes
+    hashes.add(hash)

tests/test_results_clear.py

@@ -0,0 +1,127 @@
+import asyncio
+import base64
+import json
+from pathlib import Path
+
+import pytest
+import pytest_asyncio
+from asgi_lifespan import LifespanManager
+from httpx import ASGITransport, AsyncClient
+
+from docling_serve.app import create_app
+from docling_serve.settings import docling_serve_settings
+
+
+@pytest.fixture(scope="session")
+def event_loop():
+    return asyncio.get_event_loop()
+
+
+@pytest_asyncio.fixture(scope="session")
+async def app():
+    app = create_app()
+
+    async with LifespanManager(app) as manager:
+        print("Launching lifespan of app.")
+        yield manager.app
+
+
+@pytest_asyncio.fixture(scope="session")
+async def client(app):
+    async with AsyncClient(
+        transport=ASGITransport(app=app), base_url="http://app.io"
+    ) as client:
+        print("Client is ready")
+        yield client
+
+
+async def convert_file(client: AsyncClient):
+    doc_filename = Path("tests/2408.09869v5.pdf")
+    encoded_doc = base64.b64encode(doc_filename.read_bytes()).decode()
+
+    payload = {
+        "options": {
+            "to_formats": ["json"],
+        },
+        "file_sources": [{"base64_string": encoded_doc, "filename": doc_filename.name}],
+    }
+
+    response = await client.post("/v1alpha/convert/source/async", json=payload)
+    assert response.status_code == 200, "Response should be 200 OK"
+
+    task = response.json()
+
+    print(json.dumps(task, indent=2))
+
+    while task["task_status"] not in ("success", "failure"):
+        response = await client.get(f"/v1alpha/status/poll/{task['task_id']}")
+        assert response.status_code == 200, "Response should be 200 OK"
+        task = response.json()
+        print(f"{task['task_status']=}")
+        print(f"{task['task_position']=}")
+
+        await asyncio.sleep(2)
+
+    assert task["task_status"] == "success"
+
+    return task
+
+
+@pytest.mark.asyncio
+async def test_clear_results(client: AsyncClient):
+    """Test removal of task."""
+
+    # Set long delay deletion
+    docling_serve_settings.result_removal_delay = 100
+
+    # Convert and wait for completion
+    task = await convert_file(client)
+
+    # Get result once
+    result_response = await client.get(f"/v1alpha/result/{task['task_id']}")
+    assert result_response.status_code == 200, "Response should be 200 OK"
+    print("Result 1 ok.")
+    result = result_response.json()
+    assert result["document"]["json_content"]["schema_name"] == "DoclingDocument"
+
+    # Get result twice
+    result_response = await client.get(f"/v1alpha/result/{task['task_id']}")
+    assert result_response.status_code == 200, "Response should be 200 OK"
+    print("Result 2 ok.")
+    result = result_response.json()
+    assert result["document"]["json_content"]["schema_name"] == "DoclingDocument"
+
+    # Clear
+    clear_response = await client.get("/v1alpha/clear/results?older_then=0")
+    assert clear_response.status_code == 200, "Response should be 200 OK"
+    print("Clear ok.")
+
+    # Get deleted result
+    result_response = await client.get(f"/v1alpha/result/{task['task_id']}")
+    assert result_response.status_code == 404, "Response should be removed"
+    print("Result was no longer found.")
+
+
+@pytest.mark.asyncio
+async def test_delay_remove(client: AsyncClient):
+    """Test automatic removal of task with delay."""
+
+    # Set short delay deletion
+    docling_serve_settings.result_removal_delay = 5
+
+    # Convert and wait for completion
+    task = await convert_file(client)
+
+    # Get result once
+    result_response = await client.get(f"/v1alpha/result/{task['task_id']}")
+    assert result_response.status_code == 200, "Response should be 200 OK"
+    print("Result ok.")
+    result = result_response.json()
+    assert result["document"]["json_content"]["schema_name"] == "DoclingDocument"
+
+    print("Sleeping to wait the automatic task deletion.")
+    await asyncio.sleep(10)
+
+    # Get deleted result
+    result_response = await client.get(f"/v1alpha/result/{task['task_id']}")
+    assert result_response.status_code == 404, "Response should be removed"