auto-rag

Need vectorstores testing for all except faiss

application/api/answer/routes.py

@@ -446,7 +446,8 @@ class Stream(Resource):
             attachment_ids = data.get("attachments", [])
 
             index = data.get("index", None)
-            chunks = int(data.get("chunks", 2))
+            chunks_from_request = data.get("chunks", 2)
+            chunks = chunks_from_request if str(chunks_from_request) == 'Auto' else int(chunks_from_request)
             token_limit = data.get("token_limit", settings.DEFAULT_MAX_HISTORY)
             retriever_name = data.get("retriever", "classic")
             agent_id = data.get("agent_id", None)
@@ -620,7 +621,8 @@ class Answer(Resource):
             )
             conversation_id = data.get("conversation_id")
             prompt_id = data.get("prompt_id", "default")
-            chunks = int(data.get("chunks", 2))
+            chunks_from_request = data.get("chunks", 2)
+            chunks = chunks_from_request if str(chunks_from_request) == 'Auto' else int(chunks_from_request)
             token_limit = data.get("token_limit", settings.DEFAULT_MAX_HISTORY)
             retriever_name = data.get("retriever", "classic")
             agent_type = settings.AGENT_NAME
@@ -814,7 +816,8 @@ class Search(Resource):
 
         try:
             question = data["question"]
-            chunks = int(data.get("chunks", 2))
+            chunks_from_request = data.get("chunks", 2)
+            chunks = chunks_from_request if str(chunks_from_request) == 'Auto' else int(chunks_from_request)
             token_limit = data.get("token_limit", settings.DEFAULT_MAX_HISTORY)
             retriever_name = data.get("retriever", "classic")
 

application/retriever/classic_rag.py

@@ -2,11 +2,16 @@ import logging
 from application.core.settings import settings
 from application.llm.llm_creator import LLMCreator
 from application.retriever.base import BaseRetriever
-
 from application.vectorstore.vector_creator import VectorCreator
 
+logger = logging.getLogger(__name__)
 
 class ClassicRAG(BaseRetriever):
+    # Settings for Auto-Chunking
+    AUTO_CHUNK_MIN: int = 0
+    AUTO_CHUNK_MAX: int = 10
+    SIMILARITY_SCORE_THRESHOLD: float = 0.5
+    
     def __init__(
         self,
         source,
@@ -47,6 +52,7 @@ class ClassicRAG(BaseRetriever):
         self.question = self._rephrase_query()
         self.vectorstore = source["active_docs"] if "active_docs" in source else None
         self.decoded_token = decoded_token
+        self.actual_chunks_retrieved = 0
 
     def _rephrase_query(self):
         if (
@@ -77,8 +83,66 @@ class ClassicRAG(BaseRetriever):
             return self.original_question
 
     def _get_data(self):
+        if self.chunks == 'Auto':
+            return self._get_data_auto()
+        else:
+            return self._get_data_classic()
+
+    def _get_data_auto(self):
+        if not self.vectorstore:
+            self.actual_chunks_retrieved = 0
+            return []
+
+        docsearch = VectorCreator.create_vectorstore(
+            settings.VECTOR_STORE, self.vectorstore, settings.EMBEDDINGS_KEY
+        )
+        
+        try:
+            docs_with_scores = docsearch.search_with_scores(self.question, k=self.AUTO_CHUNK_MAX)
+        except Exception as e:
+            logger.error(f"Error during search_with_scores: {e}", exc_info=True)
+            self.actual_chunks_retrieved = 0
+            return []
+        
+        if not docs_with_scores:
+            self.actual_chunks_retrieved = 0
+            return []
+
+        candidate_docs = []
+        for doc, score in docs_with_scores:
+            if score >= self.SIMILARITY_SCORE_THRESHOLD:
+                candidate_docs.append(doc)
+                
+        if len(candidate_docs) < self.AUTO_CHUNK_MIN and self.AUTO_CHUNK_MIN > 0:
+            final_docs_to_format = [doc for doc, score in docs_with_scores[:self.AUTO_CHUNK_MIN]]
+        else:
+            final_docs_to_format = candidate_docs
+            
+        self.actual_chunks_retrieved = len(final_docs_to_format)
+        
+        if not final_docs_to_format:
+            return []
+
+        formatted_docs = [
+            {
+                "title": i.metadata.get(
+                    "title", i.metadata.get("post_title", i.page_content)
+                ).split("/")[-1],
+                "text": i.page_content,
+                "source": (
+                    i.metadata.get("source")
+                    if i.metadata.get("source")
+                    else "local"
+                ),
+            }
+            for i in final_docs_to_format
+        ]
+        logger.info(f"AutoRAG: Retrieved {self.actual_chunks_retrieved} chunks for query '{self.original_question}'.")
+        return formatted_docs
+
+    def _get_data_classic(self):
         if self.chunks == 0:
-            docs = []
+            return []
         else:
             docsearch = VectorCreator.create_vectorstore(
                 settings.VECTOR_STORE, self.vectorstore, settings.EMBEDDINGS_KEY
@@ -98,8 +162,7 @@ class ClassicRAG(BaseRetriever):
                 }
                 for i in docs_temp
             ]
-
+            return docs
-        return docs
 
     def gen():
         pass
@@ -111,12 +174,24 @@ class ClassicRAG(BaseRetriever):
         return self._get_data()
 
     def get_params(self):
-        return {
+        params = {
             "question": self.original_question,
             "rephrased_question": self.question,
             "source": self.vectorstore,
-            "chunks": self.chunks,
             "token_limit": self.token_limit,
             "gpt_model": self.gpt_model,
             "user_api_key": self.user_api_key,
         }
+        if self.chunks == 'Auto':
+            params.update({
+                "chunks_mode": "Auto",
+                "chunks_retrieved_auto": self.actual_chunks_retrieved,
+                "auto_chunk_min_setting": self.AUTO_CHUNK_MIN,
+                "auto_chunk_max_setting": self.AUTO_CHUNK_MAX,
+                "similarity_threshold_setting": self.SIMILARITY_SCORE_THRESHOLD,
+            })
+        else:
+            params["chunks_mode"] = "Classic"
+            params["chunks"] = self.chunks
+
+        return params

application/retriever/retriever_creator.py

@@ -2,7 +2,6 @@ from application.retriever.classic_rag import ClassicRAG
 from application.retriever.duckduck_search import DuckDuckSearch
 from application.retriever.brave_search import BraveRetSearch
 
-
 class RetrieverCreator:
     retrievers = {
         "classic": ClassicRAG,

application/vectorstore/base.py

@@ -58,6 +58,10 @@ class BaseVectorStore(ABC):
     def search(self, *args, **kwargs):
         pass
 
+    @abstractmethod
+    def search_with_scores(self, query: str, k: int, *args, **kwargs):
+        pass
+
     def is_azure_configured(self):
         return settings.OPENAI_API_BASE and settings.OPENAI_API_VERSION and settings.AZURE_DEPLOYMENT_NAME
 

application/vectorstore/elasticsearch.py

@@ -109,6 +109,46 @@ class ElasticsearchStore(BaseVectorStore):
             doc_list.append(Document(page_content = hit['_source']['text'], metadata = hit['_source']['metadata']))
         return doc_list
     
+    def search_with_scores(self, query: str, k: int, *args, **kwargs):
+        embeddings = self._get_embeddings(settings.EMBEDDINGS_NAME, self.embeddings_key)
+        vector = embeddings.embed_query(query)
+        knn = {
+            "filter": [{"match": {"metadata.source_id.keyword": self.source_id}}],
+            "field": "vector",
+            "k": k,
+            "num_candidates": 100,
+            "query_vector": vector,
+        }
+        full_query = {
+            "knn": knn,
+            "query": {
+                "bool": {
+                    "must": [
+                        {
+                            "match": {
+                                "text": {
+                                    "query": question,
+                                }
+                            }
+                        }
+                    ],
+                    "filter": [{"match": {"metadata.source_id.keyword": self.source_id}}],
+                }
+            },
+            "rank": {"rrf": {}},
+        }
+        resp = self.docsearch.search(index=self.index_name, query=full_query['query'], size=k, knn=full_query['knn'])
+
+        docs_with_scores = []
+        for hit in resp['hits']['hits']:
+            score = hit['_score']
+            # Normalize the score. Elasticsearch returns a score of 1.0 + cosine similarity.
+            similarity = max(0, score - 1.0)
+            doc = Document(page_content=hit['_source']['text'], metadata=hit['_source']['metadata'])
+            docs_with_scores.append((doc, similarity))
+
+        return docs_with_scores
+
     def _create_index_if_not_exists(
             self, index_name, dims_length
         ):

application/vectorstore/faiss.py

@@ -63,6 +63,18 @@ class FaissStore(BaseVectorStore):
     def search(self, *args, **kwargs):
         return self.docsearch.similarity_search(*args, **kwargs)
     
+    def search_with_scores(self, query: str, k: int, *args, **kwargs):
+        docs_and_distances = self.docsearch.similarity_search_with_score(query, k, *args, **kwargs)
+        
+        # Convert L2 distance to a normalized similarity score (0-1, higher is better)
+        docs_and_similarities = []
+        for doc, distance in docs_and_distances:
+            if distance < 0: distance = 0
+            similarity = 1 / (1 + distance)
+            docs_and_similarities.append((doc, similarity))
+            
+        return docs_and_similarities
+
     def add_texts(self, *args, **kwargs):
         return self.docsearch.add_texts(*args, **kwargs)
 

application/vectorstore/lancedb.py

@@ -2,6 +2,8 @@ from typing import List, Optional
 import importlib
 from application.vectorstore.base import BaseVectorStore
 from application.core.settings import settings
+from application.vectorstore.document_class import Document
+
 
 class LanceDBVectorStore(BaseVectorStore):
     """Class for LanceDB Vector Store integration."""
@@ -87,6 +89,23 @@ class LanceDBVectorStore(BaseVectorStore):
         results = self.docsearch.search(query_embedding).limit(k).to_list()
         return [(result["_distance"], result["text"], result["metadata"]) for result in results]
 
+    def search_with_scores(self, query: str, k: int, *args, **kwargs):
+        """Perform a similarity search with scores."""
+        self.ensure_table_exists()
+        query_embedding = self._get_embeddings(settings.EMBEDDINGS_NAME, self.embeddings_key).embed_query(query)
+        results = self.docsearch.search(query_embedding).limit(k).to_list()
+
+        docs_with_scores = []
+        for result in results:
+            distance = result.get('_distance', float('inf'))
+            if distance < 0: distance = 0
+            # Convert L2 distance to a normalized similarity score
+            similarity = 1 / (1 + distance)
+            doc = Document(page_content=result['text'], metadata=result["metadata"])
+            docs_with_scores.append((doc, similarity))
+            
+        return docs_with_scores
+    
     def delete_index(self):
         """Delete the entire LanceDB index (table)."""
         if self.table:

application/vectorstore/milvus.py

@@ -26,6 +26,16 @@ class MilvusStore(BaseVectorStore):
         expr = f"source_id == '{self._source_id}'"
         return self._docsearch.similarity_search(query=question, k=k, expr=expr, *args, **kwargs)
     
+    def search_with_scores(self, query: str, k: int, *args, **kwargs):
+        expr = f"source_id == '{self._source_id}'"
+        docs_and_distances = self._docsearch.similarity_search_with_score(query, k, expr=expr, *args, **kwargs)
+        docs_with_scores = []
+        for doc, distance in docs_and_distances:
+            similarity = 1.0 - distance
+            docs_with_scores.append((doc, max(0, similarity)))
+        
+        return docs_with_scores
+
     def add_texts(self, texts: List[str], metadatas: Optional[List[dict]], *args, **kwargs):
         ids = [str(uuid4()) for _ in range(len(texts))]
 

application/vectorstore/mongodb.py

@@ -63,6 +63,40 @@ class MongoDBVectorStore(BaseVectorStore):
             results.append(Document(text, metadata))
         return results
     
+    def search_with_scores(self, query: str, k: int, *args, **kwargs):
+        query_vector = self._embedding.embed_query(query)
+
+        pipeline = [
+            {
+                "$vectorSearch": {
+                    "queryVector": query_vector,
+                    "path": self._embedding_key,
+                    "limit": k,
+                    "numCandidates": k * 10,
+                    "index": self._index_name,
+                    "filter": {"source_id": {"$eq": self._source_id}},
+                }
+            },
+            {
+                "$addFields": {
+                    "score": {"$meta": "vectorSearchScore"}
+                }
+            }
+        ]
+
+        cursor = self._collection.aggregate(pipeline)
+
+        results = []
+        for doc in cursor:
+            score = doc.pop("score", 0.0)
+            text = doc.pop(self._text_key)
+            doc.pop("_id")
+            doc.pop(self._embedding_key, None)
+            metadata = doc
+            doc = Document(page_content=text, metadata=metadata)
+            results.append((doc, score))
+        return results
+
     def _insert_texts(self, texts, metadatas):
         if not texts:
             return []

application/vectorstore/qdrant.py

@@ -36,6 +36,9 @@ class QdrantStore(BaseVectorStore):
     def search(self, *args, **kwargs):
         return self._docsearch.similarity_search(filter=self._filter, *args, **kwargs)
     
+    def search_with_scores(self, query: str, k: int, *args, **kwargs):
+        return self._docsearch.similarity_search_with_score(query=query, k=k, filter=self._filter, *args, **kwargs)
+
     def add_texts(self, *args, **kwargs):
         return self._docsearch.add_texts(*args, **kwargs)
 

frontend/src/settings/General.tsx

@@ -36,7 +36,7 @@ export default function General() {
     { label: '繁體中文（臺灣）', value: 'zhTW' },
     { label: 'Русский', value: 'ru' },
   ];
-  const chunks = ['0', '2', '4', '6', '8', '10'];
+  const chunks = ['Auto', '0', '2', '4', '6', '8', '10'];
   const token_limits = new Map([
     [0, t('settings.general.none')],
     [100, t('settings.general.low')],