Comparisons · Last updated 2026-09-21

Jev vs retrieval-only search

Retrieval-only search returns the nearest chunks. It does not say whether a chunk answers the question, contradicts it, or smuggles a prompt injection. Jev is the cheap structured judge you run after (or instead of) showing the top-k.

Unofficial comparison. Official patterns: rerank, semantic find. We do not invent recall@10. Not TypeSafe official.

Criteria

What retrieval already does well

• Keyword / lexical match (BM25)
• Dense similarity when you have embeddings
• Fast shortlists (30-passage lists in TypeSafe’s rerank writeup)

What it does not do: calibrated “this line is the answer,” “no answer exists,” or “this chunk is hostile.”

What to add

1. Retriever → Noul/Score per candidate (parallel questions on one request if the shortlist is small enough).
2. Code sorts by noul or score; drops below τ.
3. A companion Noul: “Does the document contain an answer at all?” — official line-by-line pattern. If that Noul is low, do not show a best-effort chunk as if it were a hit.
4. Never dump the whole corpus into state (jaggedness + token bill).

Cite vs generate: retrieval-only UIs should display passages, not a hallucinated synthesis. If you add an LLM writer, add a citation check.

When retrieval-only still wins

• Intranet search where users want links, not judgments
• You cannot send document text to any model
• The index already is the product

When not to use Jev

Do not replace the index. Jev will not ANN-search a million embeddings. Official context envelopes are vendor-stated (64k request / 32k state+longest on the models page) — a corpus does not fit.

Do not skip the “none exists” question and then proudly show chunk #1. A high cosine to a near-miss is still a miss.

Do not call Jev on every tile of a browse page. That is how you buy 429s. Judge the shortlist you already paid to retrieve.

Do not ask Jev to write the answer from the hits. Generation is a jaggedness failure mode. Show passages, or hand a filtered set to an LLM and citation-check the claims.

Pairing-specific failure modes

• Shortlist of distractors. jev-1.13 loses accuracy as irrelevant state grows. Filter boilerplate before the judge.
• Math in the query. “Which passage has the later date?” — extract dates in code; Jev treats dates as text.
• Injection in a retrieved chunk. State is not hostile by default. A chunk that says “ignore the query” can move the Noul. Test a hostile slice.
• Schema-safe top-1. A well-typed “best” label is not a correct answer. Use confidence and the existence Noul.

Pin jev-1.13.0 after you fit τ. TypeSafe’s published list price for jev-1.13 is $0.042 per million input tokens (vendor claim — confirm on the models page); output tokens are free. Unused distractors still bill as input.

Production sketch

1. BM25 or dense retrieve K (keep K small; TypeSafe’s rerank writeup uses shortlists, not the warehouse).
2. Strip nav chrome. One Noul or Score per {query, passage} plus one existence Noul on the document.
3. Code drops below τ or when existence is low. Sort the rest. Do not invent recall@10 here — measure on your corpus.
4. UI returns passage ids and spans. If an LLM writes a sentence, run the citation pair check before display.

Hub: Comparisons. Siblings: RAG pipelines, passage classification.

FAQ

Is Jev a reranker checkpoint? No. You pay per evaluation (vendor: input tokens). It is a request-time judge, not a trained ColBERT.

Can I skip retrieval if the wiki is “small”? Not if “small” still blows the state budget or is mostly distractors. Retrieve first.

Does Jev browse my intranet? No. Your retriever feeds text.

Disclaimer

This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.