UPDATE: Hybrid Search

functions.py

@@ -2,6 +2,7 @@ from langchain.retrievers.contextual_compression import ContextualCompressionRet
 from langchain_core.runnables import RunnablePassthrough, RunnableLambda
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_qdrant import QdrantVectorStore
+from langchain_qdrant import RetrievalMode
 from langchain_core.prompts.chat import ChatPromptTemplate
 from langchain_core.output_parsers import StrOutputParser
 from langchain.retrievers import ParentDocumentRetriever
@@ -12,6 +13,7 @@ from langchain.storage import InMemoryStore
 from langchain.docstore.document import Document
 from langchain_huggingface import HuggingFaceEmbeddings
 from langchain.retrievers import ContextualCompressionRetriever
+from langchain_qdrant import FastEmbedSparse
 from langchain.retrievers.document_compressors import FlashrankRerank
 from supabase.client import create_client
 from qdrant_client import QdrantClient
@@ -25,11 +27,12 @@ client = create_client(os.environ["SUPABASE_URL"], os.environ["SUPABASE_KEY"])
 qdrantClient = QdrantClient(url=os.environ["QDRANT_URL"], api_key=os.environ["QDRANT_API_KEY"])
 model_kwargs = {"device": "cuda"}
 encode_kwargs = {"normalize_embeddings": True}
-embeddings = HuggingFaceEmbeddings(
+vectorEmbeddings = HuggingFaceEmbeddings(
     model_name = "BAAI/bge-m3",
     model_kwargs = model_kwargs,
     encode_kwargs = encode_kwargs
 )
+sparseEmbeddings = FastEmbedSparse(model = "Qdrant/BM25", threads = 100, parallel = 0)
 prompt = """
 ### Role
     - **Primary Function**: You are an AI chatbot dedicated to assisting users with their inquiries, issues, and requests. Your goal is to deliver excellent, friendly, and efficient responses at all times. Listen attentively, understand user needs, and provide the best assistance possible or direct them to appropriate resources. If a question is unclear, ask for clarification. Always conclude your replies on a positive note.
@@ -116,7 +119,8 @@ def createTable(tablename: str):
 
 def addDocuments(text: str, vectorstore: str):
     try:
-        global embeddings
+        global vectorEmbeddings
+        global sparseEmbeddings
         global store
         parentSplitter = RecursiveCharacterTextSplitter(
             chunk_size = 2100,
@@ -128,10 +132,12 @@ def addDocuments(text: str, vectorstore: str):
         )
         texts = [Document(page_content = text)]
         vectorstore = QdrantVectorStore.from_existing_collection(
-            embedding = embeddings,
+            embedding = vectorEmbeddings,
+            sparse_embedding=sparseEmbeddings,
             collection_name=vectorstore,
             url=os.environ["QDRANT_URL"],
-            api_key=os.environ["QDRANT_API_KEY"]
+            api_key=os.environ["QDRANT_API_KEY"],
+            retrieval_mode=RetrievalMode.HYBRID
         )
         retriever = ParentDocumentRetriever(
             vectorstore=vectorstore,
@@ -178,13 +184,16 @@ def trimMessages(chain_input):
 def answerQuery(query: str, vectorstore: str, llmModel: str = "llama3-70b-8192") -> str:
     global prompt 
     global client
-    global embeddings
+    global vectorEmbeddings
+    global sparseEmbeddings
     vectorStoreName = vectorstore
     vectorstore = QdrantVectorStore.from_existing_collection(
-        embedding = embeddings,
+        embedding = vectorEmbeddings,
+        sparse_embedding=sparseEmbeddings,
         collection_name=vectorstore,
         url=os.environ["QDRANT_URL"],
-        api_key=os.environ["QDRANT_API_KEY"]
+        api_key=os.environ["QDRANT_API_KEY"],
+        retrieval_mode=RetrievalMode.HYBRID
     )
     retriever = ParentDocumentRetriever(
         vectorstore=vectorstore,

requirements.txt

@@ -1,5 +1,6 @@
 huggingface-hub
 fastapi
+fastembed-gpu
 flashrank
 gradio
 langchain