customerChatbotLanggraph

A chatbot with a LangGraph workflow that incorporates query rewriting RAG using your documentation, only structured output to reduce hallucination, rewording of unhelpful responses, but no endless loops, detection when manual intervention is necessary, and integration of issue tracking.

The LangGraph graph

The nodes are:

• classify_expectation classifies whether the customer expects a response (if not, go directly to classify_issue_tracker_item)
• classify_issue_tracker_item classifies whether a new issue tracker item needs to be created
• rewrite_query rewrites the query TOP_K_FUSION_QUERIES times for better chunk coverage in RAG
• get_chunks gets the TOP_K_CHUNKS chunks for the original query and all rewritten queries, each with its top similarity score
• classify_missing_info classifies whether there is information missing from the customer.
• draft_response creates a draft response to the customer
• classify_helpfulness classifies whether the draft response moves the chat forward towards a solution by giving or asking for relevant information (if not, do a classify_manual_intervention)
• classify_manual_intervention classifies whether a human should intervene (if not and the number of retries for the current customer query is < RETRIES, go back to draft_response).