GraphRAG in Production

This course contains the use of artificial intelligence.

The voice-over in this course is synthesized with a text-to-speech model from scripts written and reviewed by the instructor, and the on-screen material (notebooks, code, slides) is the instructor's own work.

A tutorial corpus never taught anyone what production feels like. This course runs GraphRAG on real public documents — US congressional records, bills and committee reports about artificial intelligence legislation, all public domain — and answers the question that decides whether you ship it: is this graph actually any good?

Section 1 is the document work nobody demos: downloading from govinfo, parsing PDFs and HTML, headers and reprinted text, a token diet that cut our corpus by a quarter, and a cost estimate before you spend anything. Then a production index, with settings, entity types and prompt tuning, and its real wall time and bill. Section 2 evaluates it: graph-structure metrics, entity-extraction quality judged by an LLM against the source text, automatically generated evaluation questions, and answer quality for local search, global search and a plain vector baseline. Section 3 builds multi-source graphs with provenance so an answer can be traced back to the source type it came from. Section 4 puts a function-calling agent in front of the graph and compares it with a simple router.

Every number in the course comes from a run you can repeat — including the uncomfortable one: on simple factual questions the plain vector baseline outscored both graph retrievals, at about a third of the cost of local search and a thirtieth of the cost of global search. That is taught as the first decision rule, not hidden. The finished indexes ship with the course, so you can follow every notebook without paying to re-index.