I’ve been working on a chatbot using the ConversationalRetrievalChain with a simple vector store (Chroma) and OpenAI. The first few user questions work perfectly—it pulls relevant snippets and gives a coherent answer. But after about three or four turns in the same conversation, the response starts to degrade into unrelated text or outright gibberish.
I’m using the standard `chain = ConversationalRetrievalChain.from_llm(...)` setup with `return_source_documents=True`. The chat history is managed via `chain({"question": current_question, "chat_history": chat_history})`. Has anyone else hit this wall? I’m wondering if the issue is in how the history is being formatted or if the chain is somehow corrupting the memory buffer. Any pointers would be really helpful.