Recomposing Context from Relevant Strips

Where we left off

Lesson 10 graded weather-station.md's eight sentences individually and found exactly one relevant. This lesson does something with that information: instead of handing generation the whole 685-character chunk (Beginner tier's keep-or-drop-the-whole-chunk choice), recompose a new, shorter context out of only the strips that graded relevant. This is the paper's knowledge refinement, the action taken on a chunk graded "correct."

Decompose, then recompose

That's the paper's own phrase for this: decompose a chunk into strips (Lesson 10), grade each one, then recompose the ones that survived back into a single piece of text, in their original order. It's the same idea as Lesson 5's filtering, applied one level down, filtering sentences within a kept chunk instead of filtering whole chunks within a retrieved set.

The code, piece by piece

def recompose(question: str, chunk_text: str) -> str:
    strips = split_into_strips(chunk_text)
    relevant_strips = [s for s in strips if grade_strip(question, s) == "relevant"]
    return " ".join(relevant_strips)

Split, grade, filter, and glue back together, four ideas you've already seen (Lessons 5 and 10), combined into one function. The result isn't a summary or a paraphrase, it's a subset of the original sentences, verbatim, so nothing generation reads was invented by this step.