Grounded RAG — Testing skill for Claude Code

Grounded-RAG — answers anchored in clinical guidelines

   

A retrieval-augmented assistant for doctors that answers questions about a corpus of **clinical guidelines** and backs every clinical claim with a **literal, verified quote** from them. Built on LangGraph with five interchangeable retrieval architectures (hybrid search, self-ask, LightRAG, PathRAG, HippoRAG 2), an LLM judge that can reject its own answer, and a citation layer that refuses anything it cannot find in the source text.

Demonstrated on the **HIV guidelines from GeSIDA** (seven documents, in Spanish). The engine is domain-agnostic: the guidelines, the vocabulary and the prompts' clinical content live in configuration files, not in the code.


**Status:** prototype for demonstration, no real users. Clinical decision-support tool: it does not replace medical judgement.

What it does

• Cites literally. The model must quote the sentence that supports each claim; evidence.py checks every quote against the chunk it cites and shows the section without a quote rather than a citation the guidelines do not back. Evidence grades (A-I…C-III) ride along.
• Validates before showing. A judge (gpt-4o) checks relevance and grounding; a rejected answer triggers a re-retrieval aimed at the unsupported claims, and an exhausted retry ends in a safe message, never in an unverified answer.
• Retrieves five ways. One contract, five architectures, chosen at runtime; the four graph modes share the same tail so an A/B measures selection and nothing else.
• Remembers the patient. Facts mentioned in one question steer re