grag
Ask a pile of documents a question, and get the answer with the passage it came from.
What happens when a client changes their risk profile?
An answer in two sentences, each with its source:21
The job
- When
- the answer is buried somewhere in a few hundred pages of product documentation,
- I want
- to ask in plain language and see the source,
- So
- I can trust the answer and get on with the work.
What it does
grag is a set of small retrieval services built on Gemini’s File Search API. Documents go into a store once. After that, a question comes back as an answer with citations, so each claim can be checked against the page it came from.
I use it for product and business knowledge: the material a product person needs daily and can never find twice.
How it is made
The interesting decision was what not to build. Chunking, embedding and indexing are handled by the managed service, so the code is mostly about what goes into each store, how it is kept current, and how the answers are presented.
Worth noticing
It is the project here I can show the least of. The value is in the documents, and those stay where they are.