Implement Corrective RAG (Web Search Fallback)

[joelabreurojas]

Goal

To make the ProVAI agent more robust by implementing Corrective-RAG (CRAG). This involves adding a web search tool as a fallback mechanism when the initial document retrieval fails to find relevant information.

To-Do List

• Pre-requisite: The LangGraph state machine from Issue Plan and Execute DB Migration to PostgreSQL #43 must be complete.
• Add Dependency: Add tavily-python to pyproject.toml.
• Create Web Search Tool:
  • Create a WebSearchService that uses the Tavily API to perform a search and return the results as LangChain Document objects.
  • Add a web_search node to your LangGraph (graph_nodes.py).
• Implement Document Grader Node:
  • Create a grade_documents(state) node that uses an LLM to determine if the retrieved documents are relevant to the user's question. This grader should return a simple "yes" or "no".
• Update Graph Logic:
  • Add the new grade_documents and web_search nodes to the graph.
  • Modify the conditional edge after the retrieve node. If the grade_documents node returns "no", the edge should route to the web_search node. If "yes", it should route to the generate node.
• Write tests for cases where web search should be triggered.

Acceptance Criteria

• When a query is asked for which there is no relevant information in the local documents, the agent correctly identifies this, performs a web search, and uses the web results to generate an answer.
• The decision-making process is clearly visible in the LangSmith trace.