How can I use chat history for context-aware semantic search with LangChain and Pinecone?

Below is one approach using LangChain that can help you handle chat history and semantic search together. The idea is to:

1. Maintain a chat history list that stores the previous conversation (both user and AI messages).
2. Pass this chat history along with the user's current query to your retrieval chain.
3. Optionally, use a query transformation chain to rephrase follow-up questions using context from the chat history.

For example, you can implement it as follows:

PYTHONCopied
from langchain_core.messages import AIMessage, HumanMessage

# Initialize an empty chat history
chat_history = []

# Initial query (e.g., "What is PHP?")
initial_question = "What is PHP?"
ai_response_1 = rag_chain.invoke({"input": initial_question, "chat_history": chat_history})

# Update the chat history after receiving an answer
chat_history.extend([
    HumanMessage(content=initial_question),
    AIMessage(content=ai_response_1["answer"]),
])

# Now, a follow-up query like "How to use it?" will have context.
follow_up_question = "How to use it?"
ai_response_2 = rag_chain.invoke({"input": follow_up_question, "chat_history": chat_history})

print(ai_response_2["answer"])

In this example, the follow-up question is processed along with the chat history. This way, the system is aware that "it" refers to "PHP" based on the previous conversation.

Additionally, you could build a sub-chain that reformulates the user query using the full conversation history before sending it to the retriever. In this design:

• Combine the current query and chat history into a single input.
• Pass that input through an LLM (or a dedicated chain) to generate a rephrased query.
• Use the rephrased query to perform semantic search against the Pinecone vector database.

This approach ensures your chatbot retrieves documents considering the context provided by previous messages, ultimately improving the relevance of the response.