Ship the Support Bot

shipping the finished bot

Part of: FAQ Support Bot

Every piece is built: numbering, stuffing, keyword grounding, strict refusal, citation verification, and a cost ceiling. This lesson wires them into one function you'd actually run and closes out the project. The full pipeline, in order Read it top to bottom and the whole project is one straight line: check the budget once, guard the input, pre-check grounding cheaply before spending a token, call the model with the stuffed FAQ, then verify what came back before trusting it. Nothing here is new. Lesson 8 is assembly, not invention. What "done" means for this bot - It runs from a clean start: load an FAQ file, call answer question in a loop reading from stdin. - It never answers from outside the FAQ, verified twice, once by keyword pre-check and once by citation verification. - Every real answer names its line, and every refusal is the same calm string. - A too-large FAQ fails loudly at load time instead of quietly ballooning your bill. If those four hold, you've got a shippable support bot, not a demo that happens to work on the examples you tried. Where this fits, and where it doesn't This design is right for a small, stable FAQ: a handful to a few dozen entries that don't change

Challenge: Full Pipeline Trace