ChromaDB get() or query() returns empty results after embedding updates in...

distro_dave

I am building a local RAG pipeline using LangChain, ChromaDB, and a local embedding model. The document ingestion pipeline runs without errors, and the vector store initializes correctly. However, when I attempt to query the vector store immediately after updating or adding new documents, the similarity search returns empty lists [] or fails to retrieve the newly added chunks. Here is a minimal reproducible example of how I am initializing the...

Originally asked by Ashikur Rahman on Stack Exchange (stackoverflow) (CC BY-SA 4.0).

uphilltech

Great question — answering as the UphillTech team:

The issue is mostly due to the older LangChain Chroma wrapper keeping a stale collection state after add_documents() . Instead of: from langchain_community.vectorstores import Chroma use: from langchain_chroma import Chroma With newer chromadb versions, newly added embeddings are sometimes not immediately reflected in the active retriever session. Re-initializing the vector store after add_documents() usually fixes it: vector_store.add_documents(docs, ids=ids) vector_store = Chroma( collection_name="research_docs", embedding_function=embeddings, persist_directory="./chroma_db" ) Also make sure: old collections are not created with different embedding models collection count is increasing:...

Full answer by Ritik Thakur: https://stackoverflow.stackexchange.com/a/79946752 (CC BY-SA 4.0).

Need it done for your business? UphillTech.