Bounding Correction Loops

Where we left off

Lesson 16 named it as a failure mode: nothing in Lessons 6-7's rewrite loop guarantees a rewrite eventually succeeds. Left unbounded, a question the corpus genuinely can't answer (Lesson 6's France example) would keep triggering rewrite after rewrite, burning API calls forever on a question with no answer to find. This lesson adds the guard: MAX_REWRITE_ATTEMPTS, after which the pipeline gives up honestly instead of retrying indefinitely.

The code, piece by piece

MAX_REWRITE_ATTEMPTS = 2

for attempt in range(MAX_REWRITE_ATTEMPTS + 1):
    ...
    if attempt == MAX_REWRITE_ATTEMPTS:
        break

+1 because attempt 0 is the original query, not a rewrite, 2 max rewrites means 3 total tries: the original, plus two rewritten attempts. The loop exits on the final attempt without trying yet another rewrite.

attempts_note = "" if not previous_attempts[1:] else f" (previous rewrite also failed: {previous_attempts[-1]!r})"
current_query = call_model(REWRITE_PROMPT.format(question=query, previous_attempts=attempts_note)).strip()

Each rewrite is told about the previous failed attempt, so a second rewrite doesn't just paraphrase the first rewrite's wording, it's nudged toward trying a genuinely different angle instead of converging on the same dead end twice.

return (
    f"I don't have any information relevant to that question, after "
    f"{MAX_REWRITE_ATTEMPTS} rewrite attempt(s), nothing relevant was found."
)

The honest fallback, once the bound is hit, same spirit as every earlier "I don't have any information relevant" message in this course, just now explicit about how hard it tried before giving up.