Jev vs ecommerce personalization
Ecommerce personalization ranks products from behavior graphs. Jev does not store sessions or compute CF. It can check whether a generated module or merchandising claim matches policy.
Independent. Not a recsys paper. docs.typesafe.ai.
Comparison scope
“Customers who bought X” stays your ranker. “Does this banner claim a medical benefit we forbid?” is a Noul on the copy.
Criteria that decide the architecture
| Axis | Jev (System One) | Ecommerce personalization |
|---|---|---|
| Rank SKUs | No | Wins |
| Policy on copy/modules | Noul/Choice | Rules |
| Real-time browse | Too heavy if called per tile | Specialized |
| Inventory counts | Code | Code |
Decision quality and control
Calling Jev on every product tile will 429 and add latency. Precompute policy on creatives; personalize with the recsys.
Integration trade-offs
Offline: score templates and generated blurbs. Online: recsys only. Re-score when copy changes.
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
- Merchandising policy is linguistic
Prefer Ecommerce personalization when
- You need a ranked list of SKUs
Recsys ranks; Jev preflights language; rules hide OOS. See ads page for bid vs policy.
Precompute or do not call
Policy Nouls on banner copy belong in a build or CMS hook. Personalized rank stays in the recsys. Mixing them on the hot browse path is how you buy 429s with no conversion story.
What this page does not claim
-
No conversion lift claims.
-
Schema-safe output is not the same as factual correctness.
-
No independent bake-off numbers live here.
When not to use Jev
Do not replace collaborative filtering. Jev does not store sessions or rank SKUs. A schema-safe show is not a relevance score.
Do not call Jev per tile on browse. Vendor input tokens plus 429 risk will not buy conversion. Precompute.
Do not ask Jev to write product descriptions. Generation is a jaggedness failure mode. LLM writes; Jev may policy-check the blurb.
Skip Jev when the decision is price sort, inventory count, or OOS hide — those are code.
Pairing-specific failure modes
- Inventory as language. “Only 2 left” is a count in the OMS, not a Noul.
- Promo dates. “Valid through Friday” — extract and compare in code. Dates are text to
jev-1.13. - Hot-path policy. Per-request Nouls on a merch module look like personalization and show up as latency, not lift.
- Alias on the CMS hook. Pin
jev-1.13.0when you freeze policy questions;jev-latestrewrites which banners pass without a recsys change.
Production sketch
- Recsys ranks SKUs from behavior. Inventory and price sort in code.
- CMS/build: POST policy Nouls on banner copy and generated blurbs. Pin
jev-1.13.0. Store allow/review on the creative id. - Online path uses only the ranker plus the precomputed allow bit. No Jev on browse.
- Re-score when copy changes, not when a user clicks.
- Log creative id,
response.model, and whether the module rendered. No lift numbers here. Bid-vs-policy is ad bidding.
Hub: Comparisons. Siblings: vs ad bidding systems, vs lead scoring.
FAQ
Can Jev write product descriptions? No. LLM + optional Jev policy check.
Price sorting? Code.
Per-tile Jev for “personalization”? No. That is how you buy 429s. Precompute policy; rank with the recsys.
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.