Skimmer
Filter documents and result lists with Jev relevance judgments before an agent reads them.
Added to Jevfast
How it uses Jev
Skimmer turns each URL, text file, or stdin document into overlapping chunks, then sends a Noul question for each chunk: would this chunk help accomplish the caller’s goal? It retains chunks Jev scores at or above the configured threshold and also keeps chunks Jev could not judge, so a partial Jev failure does not silently drop text. The retained chunks are emitted in document order.
What Jev decides
Example document-chunk relevance question (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · item 0
YOUR APP
INSTRUCTION
Decide whether this chunk carries substance that would help accomplish the goal; ignore boilerplate, generic background, and keyword-only mentions.
STATE
Goal plus one chunk identified as item_0; optional source label and chunk line range are included when available.
JEV · NOUL
YesNo
App workflow
- Load sources
Fetch URL text, read text/PDF files, or accept nonempty stdin; record per-source load errors.
- Prepare chunks
Optionally truncate by rune count, split into overlapping chunks, and use each chunk text as one item.
- Ask Jev
Build a shared goal/context/items state with one Noul question per item; batch requests under configured estimates and retry oversized batches by halving them.
- Keep useful text
Keep answered chunks meeting the P(yes) threshold and all unjudged chunks; optionally cap kept chunks by score, then emit in source order.
Why it is interesting
It uses a goal-conditioned relevance question at chunk granularity and batches those questions under a token budget. Oversize requests are recursively split, while failed items remain available to the caller as unjudged content.