Jev vs high-recall systems
A high-recall system would rather over-include (search, safety, fraud holds). Jev will not magically maximize recall. You choose thresholds and whether other exists.
Unofficial. Pair with the high-precision page. docs.typesafe.ai.
Comparison scope
Retrieval, keyword OR-lists, and “hold if any signal” rules are recall tools. Jev is a scorer you can set low to keep more candidates for humans.
Criteria that decide the architecture
| Axis | Jev (System One) | High-recall systems |
|---|---|---|
| Miss cost vs review cost | You set τ | OR-rules / wide retrieval |
| Closed set | Choice can drop unlabeled classes unless other |
Wide nets catch unknowns |
| Calibration | Measure on your labels | Usually uncalibrated hits |
Decision quality and control
If missing a case is catastrophic, do not use a single high τ. Use Jev to rank a wide candidate set from search/rules, then humans.
Integration trade-offs
Recall layer (search/rules) → Jev Score/Noul → sort → review until budget. Do not delete the recall layer.
TypeSafe’s public models page lists jev-1.13 at $0.042 per million input tokens with output tokens free — a vendor claim, not a jev.pro measurement. Confirm on the models page before you bid.
When each approach fits
Prefer Jev when
- You have a shortlist and want a ranking / second filter
Prefer High-recall systems when
- You still need the wide net
High recall first, Jev to spend reviewer time, precision page for auto-act.
Rank the wide net, keep the net
A high-recall system would rather over-include: search, safety, fraud holds, keyword OR-lists. Jev will not magically maximize recall. You choose thresholds and whether other exists. Sketch: recall layer (search/rules) produces a wide candidate set → Jev Score/Noul on pinned jev-1.13.0 → sort → humans review until budget. Do not delete the recall layer. Pair with vs high-precision systems when you later auto-act.
Scoring a 10k-candidate dump without a shortlist is a bill even at the vendor-claimed $0.042 / M input, and a jaggedness problem (huge dirty state). If missing a case is catastrophic, do not use a single high τ. Jev does not generate extra candidates; the wide net does.
How “Jev will catch it” deletes the OR-list
- Single high threshold as the product. That is a precision move. Recall-first products keep a low τ or a human tail.
- Closed Choice without
other. Forced labels drop unlabeled classes. Unknowns still need a recall path. - Low Noul assumed to raise recall. It raises how often you treat “yes” as true. Measure misses; don’t assume.
- Calibration theater. TypeSafe trains for calibrated decisions — vendor claim. Plot reliability on your labels before you use a Score as P(hit).
- Schema-safe keep as “we didn’t miss.” Valid output is not recall@k. We publish no BM25 bake-off here.
When not to use Jev here
Do not use Jev as the wide net. Do not use it when you still need the recall layer as the product and have no shortlist. Prefer search/OR-rules/holds first. Use Jev when you have a shortlist and want a ranking or second filter so reviewers spend time on the ugly remainder.
Precision floors belong on the auto-act path, documented on the sibling high-precision page. Mixing “don’t miss” and “auto-act” in one Noul is how you silently drop the tail.
What this page does not claim
-
No recall@k vs BM25 published here.
-
Schema-safe output is not the same as factual correctness.
-
No independent bake-off numbers live here.
FAQ
Does a low Noul threshold raise recall? It raises how often you treat “yes” as true. Measure misses; don’t assume.
Is other good for recall?
It prevents forced wrong labels. Unknowns still need a recall path.
Can I drop keyword OR-lists once Jev is live? No. That is how you silently lose the long tail. Keep the net; spend Jev on the shortlist.
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.
Hub: Comparisons. Siblings: vs high precision systems, vs semantic search, uncertainty handling. Canonical: https://docs.typesafe.ai.