What this is

No new concepts in this lesson. This is the course's capstone: a small, real FastAPI service built entirely out of ideas from Lessons 1 through 24, combined into one thing. If you can read lesson.py and understand why every piece is there, you've mastered this course. If any piece feels unfamiliar, that's a sign to revisit the lesson it came from.

What it does

Ingests every fixture note into a chromadb collection at startup (Lesson 23), then serves a GET /ask endpoint that: pre-filters candidates by similarity score (Lesson 20), grades what survives (Lessons 3, 12), bounded-rewrites and re-retrieves if nothing graded relevant (Lessons 6-7, 21), falls back to external search if the bound is hit (Lesson 22), and cites its source in every answer (Lesson 15).

Where each piece came from

candidates = [c for c in chunks if c["score"] >= PRE_FILTER_MIN_SCORE]
graded = [{**c, "grade": grader.grade(query, c["text"])} for c in candidates]

Lesson 20's pre-filter, run before Lesson 3/12's real LLM grading, to cut grading calls on obviously-unrelated candidates.

while not relevant and attempts < MAX_REWRITE_ATTEMPTS:
    effective_query = call_model(REWRITE_PROMPT.format(question=effective_query)).strip()
    ...
    attempts += 1

Lesson 21's bounded rewrite loop. MAX_REWRITE_ATTEMPTS = 1 here (smaller than Lesson 21's own 2), kept low specifically so a live service doesn't spend unbounded grading calls per incoming request, this course's Lesson 19 cost lesson, applied.

if relevant:
    ...
else:
    web_result = external_search(query)

Lesson 22's real "incorrect" branch: once the bound is hit, fall through to external search (mocked here, pluggable in a real deployment) instead of giving up or looping forever.

context = "\n\n".join(f"[{c['source']}]\n{c['text']}" for c in relevant)
...
prompt = f"""...citing the source in brackets like [source.md] or [external web search]..."""

Lesson 15's citation format, extended to also credit [external web search] when that branch fires, so a user can always tell where an answer actually came from.

Try this yourself

Without looking anything up:

  • Set MAX_REWRITE_ATTEMPTS = 0 and re-run. Does the France question now go straight to external search without ever attempting an internal rewrite? Does that change anything about the Aurora question's outcome?
  • Swap mock_web_search for a function that always returns None. Confirm the France question now returns the honest "internally or externally" failure message from Lesson 21/22 instead of an answer.
  • Run uvicorn lesson:app --reload from this folder and hit GET /ask?q=... from a browser or curl, confirm it behaves identically to the TestClient calls in the script.

This is where Corrective RAG, built entirely from scratch, ends up: a small, real, citation-aware service that grades what it retrieves and corrects it, internally or externally, before ever answering. Lesson 26 is a short, code-free look at where this specific architecture still falls short, and which course in this series picks up each of those threads.