Retrieval and verification are separate failure points
Missing a file in the first-stage pool cannot be repaired by a later Noul. The README’s candidate-input option lets a caller evaluate a different recall strategy.
Inspect the supporting source ↗REPOSITORY ANALYSIS
jev-retrieval
A Rust semantic-search CLI: local BM25 generates candidates, Jev checks file windows and reranks kept results, and code emits path:line matches or JSON.
E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
- Stars
- 0
- Forks
- 0
- License
- Apache-2.0
- Last commit
- 2026-09-23
What Jev does here
Noul gates verify candidate file windows and bounded Choice reranking orders retained files; code owns candidate recall, output limits and exit status.
Adds concept-based navigation alongside literal search while preserving distinct no-match and operational-failure outcomes.
Fixed-version documentation and implementation material were inspected on 2026-10-24. No repository search or relevance benchmark was executed. Cached responses reproduce prior observations, not an independent second model trial.
What task does this project fit?
Find code or documents by a semantic description while retaining grep-shaped results.
Material analysed：
From input to action
- A local BM25 pass generates candidates unless caller-supplied candidates bypass it.
- Jev checks file windows with Noul gates and reranks kept files through bounded choices.
- Code emits ranked path:line results, optional snippets or JSON with grep-like exit statuses.
Implementation details that matter
The client retries only a bounded failure class
client.rs caches successful responses and retries transient 408/429/5xx outcomes with exponential delay. Corrupt cached records fall through to refetch instead of becoming evidence.
Inspect the supporting source ↗When to choose it
Useful for concept search inside a repository. Keep rg for exact literals and inspect both candidate recall and verified matches when a semantic query returns no result.
How to inspect it yourself
Read README’s candidates/snippets options and client.rs system_one. Trace an unparseable cached response and a non-retryable HTTP status; distinguish no matches from operational failure using the documented exit codes.
Limits of this review
- No repository search or relevance benchmark was executed.
- Cached responses reproduce prior observations, not an independent second model trial.