react · mistral ai · node.js · remix · tailwind css · shadcn

the problem

Company handbooks are long and rarely read end to end. I wanted employees to ask a question in plain words and get an answer grounded in the handbook itself, not in the model's general knowledge.

what I decided

1. 01

   retrieval before generation

   Each question is turned into an embedding with mistral-embed and matched against handbook passages stored in Supabase. Only the five closest passages above a 0.78 similarity threshold are given to the model as context.

2. 02

   answers you can scan

   The model's replies are formatted into headings, numbered steps and bullet points before they're shown, so a policy answer reads like a policy.

3. 03

   room to grow into an agent

   I also experimented with Mistral function calling, letting the model call tools such as a payment-status lookup over up to three rounds before answering.

the outcome

A chat interface with saved conversations that answers handbook questions from the handbook's own text, and a base for trying agents that use tools.