MedRAG

MedRAG runs as a LangGraph pipeline — five nodes over one shared state object, not a single call. query_analysis rewrites what you asked into something a vector index can match, retrieve embeds that with PubMedBERT and pulls the closest passages out of Supabase, and relevance_check reads them back to decide whether they're enough. Only then does synthesize write an answer. The node I care about most is relevance_check. It asks an LLM judge whether the passages actually cover the question, and if they don't it doesn't answer anyway — it routes back to query_analysis with a note about what was missing, so the next search is a different search and not the same one twice. That loop is bounded: max_query_retries caps it, and when the budget runs out MedRAG says the papers don't cover it rather than filling the gap itself.

Ask for a diagram instead of a paragraph and synthesize emits a Mermaid spec rather than prose. The catch is that a model writing Mermaid produces syntax that almost parses — a stray code fence, an unquoted label with a parenthesis in it — and one bad character renders nothing at all. So render_diagram cleans it before it ever reaches a renderer: strip the fences and preamble, quote every node label across all the shapes, then validate. The browser does the actual drawing, which keeps a headless Chrome off the server entirely.