Jev 101
Start Guides
Search
中文 EN
Retrieval heat 1 · Quiet★ 3heat incomplete (no X)

Jevector

Profile matching that encodes each profile's criteria as 64 or 256 numeric decision values with Jev or Cloudflare Clef, retrieves candidates with HNSW, checks mandatory conditions in both directions in code, then ranks by preference coverage.

Repo: lalalune/jevector ↗

What it is

Matching on criteria instead of on similar text.

What it does: A profile and its partner conditions are turned into a fixed list of decision values by asking Jev (or Clef) the same questions every time. HNSW finds candidate profiles in that vector space. Code then checks the hard conditions both ways and ranks what is left by how many preferences are met.

Why it belongs in the showcase: Uses decision outputs as features for a classical index, and compares them openly with BM25 and an embedding model.

Caveats: Small synthetic benchmark by the author; no license file.

Evidence

Public README with a benchmark on 52 synthetic profiles and eight held-out queries, measured before rule filtering: Jev 256 top-1 87.5% and MRR@10 0.938, Clef 256 top-1 100%, BM25 top-1 12.5%, BGE-small top-1 0%; the included profiles can be searched without an API key; encoding new profiles needs a TypeSafe or Cloudflare account.

Caveats

  • The author says the profiles were built around the selected criteria and the test is small; the results do not show better performance on real profiles or general search
  • No license file at intake; new repository (3 stars)
  • Running the benchmark needs both provider accounts and can incur API charges

Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.