Introduction

• Classic RAG is great for direct questions but falters in multi-turn chats. Follow-ups like “What about pricing?” often miss because there’s no memory or smart query handling.
• This agent solves that with:
  • Query Reformulation: Turns follow-ups into clear, standalone questions
  • Topic Detection: Keeps conversations within the defined domain
  • Document Vetting: Checks content quality before answering
  • Adaptive Search: Improves queries when results are weak
  • Conversation Memory: Preserves context across turns
• The example scenario is a Technology Support Knowledge Base.

System Architecture

• The workflow is a staged pipeline:
  • Query Enhancer: Reformulates using conversation history
  • Topic Validator: Confirms the question is in-domain
  • Content Retriever: Pulls relevant documents
  • Relevance Assessor: Rates document usefulness
  • Response Generator: Produces contextual answers
  • Query Optimizer: Refines searches with capped retries
• Result: a resilient multi-turn RAG that maintains quality and relevance.

Notebook & Purpose

• Notebook: See Advanced_rag_workflow.ipynb.
• Purpose: A LangGraph-powered RAG agent that reformulates questions, validates topic relevance, retrieves from a Chroma vector store, filters document relevance, and generates contextual answers.

Features

• Query Reformulation: Builds context-aware, standalone queries from conversation history.
• Topic Classification: Routes off-topic questions to a friendly handoff.
• Vector Retrieval: Uses sentence-transformers/all-MiniLM-L6-v2 + Chroma for similarity search.
• Relevance Filtering: Structured assessment to keep only useful docs.
• Response Generation: Contextual answers with conversation memory via LangGraph checkpoints.

Requirements

• Python: 3.11+
• Dependencies: Managed via pyproject.toml
  • langchain, langgraph, chromadb, sentence-transformers, langchain-openai, langchain-google-genai, langchain-huggingface, python-dotenv, jupyterlab, ipykernel

Setup

• Create and activate a virtual environment.

uv init .
uv add -r requirements.txt

• Optional: launch JupyterLab to run the notebook.

jupyter lab

Environment Variables

• Create a .env file in the project root with any keys you plan to use:
  • For Google Gemini: GOOGLE_API_KEY
  • For OpenAI: OPENAI_API_KEY
  • Optional model wiring (see next section): LLM_PROVIDER, LLM_MODEL, LLM_TEMPERATURE

Example .env:

# Provider can be google or openai
LLM_PROVIDER=openai
LLM_MODEL=gpt-5.1-codex-max
LLM_TEMPERATURE=0.1
OPENAI_API_KEY=sk-...yourkey...
# If using Google
# GOOGLE_API_KEY=AIza...yourkey...