Jev vs RAG pipelines
A RAG pipeline retrieves, then generates. Jev (TypeSafe System One) does not generate. It is a decision layer you can bolt onto retrieval: keep, drop, flag, or rerank passages — then a different model writes prose.
This unofficial comparison is for people who searched “Jev RAG.” Official cookbooks: classifying passages, rerank, citation check. Confirm methods and any published lifts there — we will not restated unofficial percentages as ours. Not affiliated; no keys sold.
Where Jev sits in the pipe
query → retrieve (BM25 / dense) → Jev per passage or pair → code filter
→ (optional) LLM generate → Jev citation check → code allow/block| Job | Jev primitive | Not Jev |
|---|---|---|
| Is this passage relevant / injected / contradictory? | Noul or Score | Writing the answer |
| Which candidate is best for this query? | Noul/Score per pair; code sorts | Replacing the retriever |
| Does the quote support the claim? | Choice + confidence | Inventing a citation |
TypeSafe’s rerank cookbook (vendor-measured): BM25 shortlists then one question per query–candidate pair; they publish top-1 / top-10 lifts on CLERC — read the cookbook for the numbers.
Failure modes (honest)
- Huge unfiltered state — official jaggedness: accuracy falls; also the prescribed reason to classify passages first.
- Asking Jev to summarize the pile — generation failure mode. Extract candidates in code or with an LLM, then Choice.
- Treating schema-safe “supports” as truth — citation check can still be wrong; use confidence to review.
- One 218-way Choice without a “none” — line-by-line search cookbook pairs a big Choice with a Noul “does the document contain an answer?”
- Adversarial passage — injected instructions can steer
jev-1.13. Test before you drop a “filter” in production.
Cite vs generate
If the product promise is “answers only from sources,” Jev can screen claims. It cannot author the sentence. Keep that boundary in the UX: show sources your code kept, not a Jev-written paragraph.
When not to use Jev
Do not replace the RAG pipeline with a single System One POST. Retrieval, chunking, and generation stay yours. Jev is a snap judge on pairs you already have.
Do not skip retrieval and dump the wiki into state. Official jaggedness plus the vendor token bill: unused distractors still count as input.
Do not ship a generated paragraph without a citation check if the UX promised grounded answers. The writer is an LLM; the judge is Jev; code is the allow/block.
Skip Jev when the product only needs links (retrieval-only) or when you cannot send passage text off-box.
When the full RAG pipeline wins
The pipeline wins whenever the user needs a paragraph. Jev will not write it. The pipeline also wins at corpus ops: embeddings, hybrid search, ACL filters. See semantic search.
Cookbook lifts stay on TypeSafe’s pages (vendor-measured). We do not invent a jev.pro RAG bake-off.
Pin jev-1.13.0 after you fit keep/drop 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.
Pairing-specific failure modes
- Generate-then-hope. An LLM writes; no pair check; a schema-safe later Choice is used as theater.
- Date/number claims. “The policy changed on 12 March” — extract the date in code; ask Jev only whether the span is about that policy.
- Filter that never drops. τ fitted on happy demos, never on contradictory or empty shortlists.
- Alias move mid-prompt. Criteria and
jev-1.13.0must version together or citation floors lie.
Production sketch
- Retrieve K. Classify or rerank each passage with pinned Jev. Code drops below τ.
- If nothing remains, return “no grounded answer” — do not generate from an empty set.
- LLM writes from the kept spans only.
- Citation-check each claim vs its span (citation services). Low confidence → withhold or review.
- Log
response.model, kept ids, and whether the writer ran. Replay after criteria edits.
Hub: Comparisons. Siblings: retrieval-only, passage classification.
FAQ
Is Jev a vector database? No. Retrieval stays yours.
Can I skip the LLM entirely? If the user only needs a routed document id or a yes/no, yes. If they need a paragraph, no.
Are cookbook top-k lifts ours? No. Vendor-measured on their pages. Re-measure on your corpus.
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.