Steer RAG chunking and index hygiene with Jev

01 · Role in the system

What Jev does here

Most RAG indexes are built with fixed-size cuts that split thoughts mid-sentence. Here a Jev call judges whether each sentence continues the point of the one before, and code places the cut inside hard token limits where the meaning actually changed; headings, tables and code blocks are handled structurally. Before anything is embedded, filler and boilerplate are dropped, and text that tries to instruct an AI is quarantined - the index never learns the poisoned passages in the first place. Chunks can be tagged against your own taxonomy with probabilities, and at query time each retrieved passage is classified as evidence, a conflict with the question, or noise, so the LLM sees only what answers. The client is a small Python package with tests, database-agnostic by design.

02 · Control boundary

Where Jev sits

Jev as the quality judge across the whole RAG lifecycle: sentence-continuation Nouls place chunk boundaries, hygiene questions quarantine planted instructions before embedding, and retrieval-time classification separates evidence from conflict and noise.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• Sentence-continuation judging adds one call per candidate boundary; very large corpora pay per chunk.
• New project (created 2026-09-30, 0 stars); no benchmark against fixed-size chunking baselines yet.
• Quarantine criteria are the author's; adversarial evaluations are not published.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• MersivMedia ↗Community · github · public · checked 2026-10-01