Module distill 

Extractive research-compression modes for prose and transcripts.

These are deterministic, heuristic distillations — no LLM in the loop — so they are cheap, reproducible, and safe to run inside a synchronous tool handler. They turn a cleaned article or transcript into the high-signal subset an agent actually needs:

• facts_scored — sentences carrying factual signals (numbers, dates, entities), each with a confidence score.
• quotes_scored — the most central / query-relevant sentences, as evidence, each with a confidence score.
• transcript_summary — de-duplicated, filler-stripped spoken text.

Functions

facts_scored

Extract sentences carrying factual signals, ranked and de-duplicated. Each sentence carries a confidence ([0.0, 1.0]) so callers can build attributable crate::core::evidence::Claims. Facts use an absolute mapping (more factual signals → higher confidence) rather than min-max, so the score is meaningful even when the top sentences tie.

quotes_scored

Extract the most central (or query-relevant) sentences as quotable evidence. Each sentence carries a source-relative confidence ([0.0, 1.0]).

split_sentences

Split text into trimmed, non-empty sentences across line boundaries.

squeeze_prose

Line-structure-preserving prose squeeze for the proxy tool-result funnel.

summarize_prose

Summarize prose to a max_chars budget, query-aware.

transcript_summary

Condense a transcript: strip filler, drop near-duplicate runs, cap length.