RAGuette

France has more than two million micro-entrepreneurs asking the same admin and tax questions, and the official answers are scattered across URSSAF, service-public.fr and impots.gouv.fr. RAGuette answers them conversationally, in French or English — and never without a citation to the official source, with the date it was last fetched.

The ingestion pipeline fetches the official pages, extracts the article text with Mozilla Readability and stores it as Markdown with its source URL and fetch date, then chunks it and embeds it into Postgres with pgvector. At question time, the closest chunks are retrieved by cosine similarity and the model is prompted to answer only from them, citing each one.

Version 0.1 works end to end: ingestion, retrieval, cited answers and a streaming chat interface, all on free tiers. The next step is an evaluation harness — question–answer pairs with a measured accuracy — because a project whose whole point is not hallucinating has to measure it.