feat: add hybrid search as retrieval mode

rag-core-api/pyproject.toml

@@ -28,6 +28,8 @@ langgraph = "^0.2.23"
 pillow = "^11.0.0"
 langchain-ollama = "^0.2.0"
 pytest-asyncio = "^0.25.0"
+langchain-community = "0.3.19"
+fastembed = "^0.6.0"
 
 [tool.poetry.group.dev.dependencies]
 debugpy = "^1.8.1"

rag-core-api/src/rag_core_api/dependency_container.py

@@ -16,7 +16,7 @@
 from langchain_community.llms.fake import FakeListLLM
 from langchain_ollama import ChatOllama
 from langchain_openai import ChatOpenAI
-from langchain_qdrant import QdrantVectorStore
+from langchain_qdrant import QdrantVectorStore, FastEmbedSparse
 from langfuse import Langfuse
 
 from rag_core_api.impl.answer_generation_chains.answer_generation_chain import (
@@ -50,6 +50,7 @@
 from rag_core_api.impl.settings.ragas_settings import RagasSettings
 from rag_core_api.impl.settings.reranker_settings import RerankerSettings
 from rag_core_api.impl.settings.retriever_settings import RetrieverSettings
+from rag_core_api.impl.settings.sparse_embedder_settings import SparseEmbedderSettings
 from rag_core_api.impl.settings.stackit_embedder_settings import StackitEmbedderSettings
 from rag_core_api.impl.settings.vector_db_settings import VectorDatabaseSettings
 from rag_core_api.impl.vector_databases.qdrant_database import QdrantDatabase
@@ -95,6 +96,7 @@ class DependencyContainer(DeclarativeContainer):
     embedder_class_type_settings = EmbedderClassTypeSettings()
     stackit_embedder_settings = StackitEmbedderSettings()
     chat_history_settings = ChatHistorySettings()
+    sparse_embedder_settings = SparseEmbedderSettings()
     chat_history_config.from_dict(chat_history_settings.model_dump())
 
     class_selector_config.from_dict(rag_class_type_settings.model_dump() | embedder_class_type_settings.model_dump())
@@ -110,22 +112,28 @@ class DependencyContainer(DeclarativeContainer):
         ),
     )
 
+    sparse_embedder = Singleton(FastEmbedSparse, **sparse_embedder_settings.model_dump())
+
     vectordb_client = Singleton(
         qdrant_client.QdrantClient,
         location=vector_database_settings.location,
     )
+
     vectorstore = Singleton(
         QdrantVectorStore,
         client=vectordb_client,
         collection_name=vector_database_settings.collection_name,
         embedding=embedder,
+        sparse_embedding=sparse_embedder,
         validate_collection_config=False,
+        retrieval_mode=vector_database_settings.retrieval_mode,
     )
 
     vector_database = Singleton(
         QdrantDatabase,
         settings=vector_database_settings,
         embedder=embedder,
+        sparse_embedder=sparse_embedder,
         vectorstore=vectorstore,
     )
 

rag-core-api/src/rag_core_api/impl/settings/sparse_embedder_settings.py

@@ -0,0 +1,23 @@
+"""Module contains settings regarding the sparse embedder."""
+
+from pydantic import Field
+from pydantic_settings import BaseSettings
+
+
+class SparseEmbedderSettings(BaseSettings):
+    """
+    Contains settings regarding the sparse embedder.
+
+    Attributes
+    ----------
+    model_name : str
+        The name of the model to be used (default "Qdrant/bm25").
+    """
+
+    class Config:
+        """Config class for reading Fields from env."""
+
+        env_prefix = "SPARSE_EMBEDDER_"
+        case_sensitive = False
+
+    model_name: str = Field(default="Qdrant/bm25")

rag-core-api/src/rag_core_api/impl/settings/vector_db_settings.py

@@ -3,6 +3,8 @@
 from pydantic_settings import BaseSettings
 from pydantic import Field
 
+from langchain_qdrant import RetrievalMode
+
 
 class VectorDatabaseSettings(BaseSettings):
     """
@@ -27,3 +29,4 @@ class Config:
     validate_collection_config: bool = Field(
         default=False
     )  # if true and collection does not exist, an error will be raised
+    retrieval_mode: RetrievalMode = Field(default=RetrievalMode.HYBRID)

rag-core-api/src/rag_core_api/impl/vector_databases/qdrant_database.py

@@ -2,7 +2,7 @@
 import logging
 
 from langchain_core.documents import Document
-from langchain_qdrant import QdrantVectorStore
+from langchain_qdrant import QdrantVectorStore, SparseEmbeddings
 from qdrant_client.http import models
 from qdrant_client.models import FieldCondition, Filter, MatchValue
 
@@ -25,6 +25,7 @@ def __init__(
         self,
         settings: VectorDatabaseSettings,
         embedder: Embedder,
+        sparse_embedder: SparseEmbeddings,
         vectorstore: QdrantVectorStore,
     ):
         """
@@ -43,6 +44,7 @@ def __init__(
             settings=settings,
             embedder=embedder,
             vectorstore=vectorstore,
+            sparse_embedder=sparse_embedder,
         )
 
     @property
@@ -167,9 +169,11 @@ def upload(self, documents: list[Document]) -> None:
         """
         self._vectorstore = self._vectorstore.from_documents(
             documents,
-            self._embedder.get_embedder(),
-            collection_name=self._settings.collection_name,
+            embedding=self._embedder.get_embedder(),
+            sparse_embedding=self._sparse_embedder,
             location=self._settings.location,
+            collection_name=self._settings.collection_name,
+            retrieval_mode=self._settings.retrieval_mode,
         )
 
     def delete(self, delete_request: dict) -> None:

rag-core-api/src/rag_core_api/vector_databases/vector_database.py

@@ -4,6 +4,7 @@
 
 from langchain_community.vectorstores import VectorStore
 from langchain_core.documents import Document
+from langchain_qdrant import SparseEmbeddings
 
 from rag_core_api.embeddings.embedder import Embedder
 from rag_core_api.impl.settings.vector_db_settings import VectorDatabaseSettings
@@ -16,6 +17,7 @@ def __init__(
         self,
         settings: VectorDatabaseSettings,
         embedder: Embedder,
+        sparse_embedder: SparseEmbeddings,
         vectorstore: VectorStore,
     ):
         """
@@ -32,6 +34,7 @@ def __init__(
         """
         self._settings = settings
         self._embedder = embedder
+        self._sparse_embedder = sparse_embedder
         self._vectorstore = vectorstore
 
     @property

rag-core-lib/pyproject.toml

@@ -12,7 +12,7 @@ packages = [{ include = "rag_core_lib", from = "src" }]
 [tool.poetry.dependencies]
 python = "^3.11"
 langchain = "^0.3.7"
-langchain-community = "0.3.7"
+langchain-community = "0.3.19"
 flashrank = "^0.2.5"
 pydantic-settings = "^2.2.1"
 pydantic = "^2.7.2"