Jev vs vector databases
A vector database stores embeddings and answers nearest-neighbor queries. Jev has no vector index. People search this pairing because RAG decks blur “retrieve” and “decide.”
Independent. Not a Pinecone/Weaviate/pgvector bake-off. Official Jev input: text state, no embeddings field. docs.typesafe.ai.
Comparison scope
Jev will not replace ANN search, metadata filters, or hybrid BM25+dense. It may sit after the DB returns chunks.
Criteria that decide the architecture
| Axis | Jev (System One) | Vector databases |
|---|---|---|
| Persist vectors | No | Yes |
| kNN / hybrid search | No | Yes |
| Keep/drop chunk | Noul/Score | Not the DB’s job |
| Mutate schema of labels | Edit criteria | Re-embed if you stored labels in vectors |
Decision quality and control
A close neighbor is not a relevant neighbor. That gap is why TypeSafe’s RAG-passage cookbook exists. Do not treat cosine as a policy decision.
Integration trade-offs
DB query → strip boilerplate → Jev questions → code filter → (optional) LLM. Watch the 32k state+longest-question budget (vendor models page).
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 need a typed decision on retrieved text
Prefer Vector databases when
- You need storage and retrieval
Vectors retrieve; Jev judges; code acts.
Judge the shortlist, do not index it
A vector database stores embeddings and answers nearest-neighbor queries. Jev has no index, no embed API, and no embeddings field on the public evaluation contract we used. Sketch: ANN or hybrid search returns k chunks → strip boilerplate → Jev Noul/Score “keep this chunk for query Q” on pinned jev-1.13.0 → code filter → optional LLM. Watch the vendor 32k state+longest-question budget on the models page.
Token list price ($0.042 / M input, output free) is a vendor claim on the models page. Sending all k=50 chunks unfiltered is how you pay for state rot and, per TypeSafe jaggedness / RAG-passage cookbook, lose accuracy. A close neighbor is not a relevant neighbor. Cosine is not a policy decision.
How RAG decks blur retrieve and decide
- Jev as the ANN store. It will not persist vectors or run kNN. If you need storage, you still buy/build the DB.
- Cosine threshold as keep/drop. That is retrieval hygiene, not a rubric. Put the rubric in Jev questions after the shortlist.
- Labels baked into vectors. If you stored yesterday’s keep/drop inside the embedding, you re-embed to change policy. Jev criteria edit without a reindex.
- Huge unfiltered state. Official reason to classify passages first. Also a bill.
- Schema-safe “relevant” as truth. A valid Noul is not “this passage supports the claim.” Citation check is a later, separate question — and still not factual proof. Jev does not generate the answer paragraph.
When not to use Jev here
Do not use Jev to replace ANN search, metadata filters, or hybrid BM25+dense. Do not use it as an embedding model. Prefer the DB alone when you only need storage and retrieval. Use Jev when you need a typed decision on retrieved text.
What this page does not claim
-
No ANN benchmark.
-
Jev is not an embedding model in the public docs we used.
-
Schema-safe output is not the same as factual correctness.
-
No independent bake-off numbers live here.
FAQ
Does Jev output embeddings? Public System One answers are typed decisions, not vectors.
Can I store Jev probabilities in the DB? Yes, as metadata your app writes — not a TypeSafe feature.
Does Jev embed the query? Not in the public docs we used. Your embedder + DB do that. Jev sees 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.
Hub: Comparisons. Siblings: vs semantic search, vs knowledge graphs, glossary retrieval grounding. Canonical: https://docs.typesafe.ai.