All labs · Lab 21 · intermediate · 45 min

Answer from your documents, with sources

Build a question answerer that reads your own documents, names the file each answer came from, and says so when the answer is not there.

You build: ingest.py and ask.py over six pages of Parcel Desk policy.

  • Python 3.11 or newer and uv
  • Ollama
  • Docker
  • Lab 20 done (the learn-pg container)
pgvectorOllama docs

  1. 01Write six short policy pages: refunds, damage, lost parcels, international, time limits, exclusions

    ls docs/*.md | wc -l

    You pass when: Six files, each under 300 words, with at least one concrete number (days, amounts).

  2. 02Split each page into chunks and store them with their file name

    uv run --with requests --with "psycopg[binary]" ingest.py docs/

    You pass when: It prints how many chunks it stored per file. Running it twice does not double the rows.

  3. 03Fetch the four closest chunks and answer only from them, naming the files

    uv run --with openai --with requests --with "psycopg[binary]" ask.py "How many days do I have to report a damaged parcel?"

    You pass when: The number from your damage page, followed by the file name it came from.

  4. 04Ask something the pages do not cover

    An answerer that guesses is worse than no answerer. This refusal is the most important line in your prompt.

    uv run --with openai --with requests --with "psycopg[binary]" ask.py "Do you ship live animals to Mars?"

    You pass when: It says the documents do not cover this. It does not make up a policy.

The check runs on the computer, not on your word: a step passes when its command exits with 0 there.

Your computer · Fleet · off

Computers are not set up on this site yet. Run each command on your own machine and compare with "You pass when".