Search returns candidates. Something then has to decide which ones matter. A reranker scores each result against the query, and a RAG filter drops retrieved passages that will not help the answer. A decision model fits both. It answers one relevance question per item and returns a probability, so you can sort by it or cut at a threshold. The examples in this hub run on Jev and other decision models, and Laya appears in some. Some search code, such as nassim-arifette/jevgrep and Bentlybro/siftr. Some rerank, such as hev's reranker and the TypeSafe AI cookbook on re-ranking. Others search personal data, like jh1373/jev-search for an Obsidian vault. Before you copy one, check how many items go into each call, what retrieves the shortlist first, whether it handles long documents, and how the author compared it with a standard reranker. Entries range from one-file command-line tools to benchmarks, so decide first whether you need something to run or something to learn from.
25 entries
Showing 25 of 25
Agent skills for Hermes, Claude Code and Codex that hand Jev the routing decisions: which model tier to use, which search results to open, which memory passages to read and which hide injected instructions. Ships a dashboard with an on, shadow and off switch and publishes measured latencies per skill.
Web search in plain language where Jev chooses sources, time ranges and search terms, then scores each result for relevance, with retrieval through Search1API. Returns links and snippets with visible relevance scores and editable filters, and writes no answers.
A grep that matches lines by meaning instead of regular expressions: Jev scores each line against a plain-language description and the tool applies a threshold, combining meanings with AND, OR and NOT.
134 starsBatch speed <1s/210 lines
Codebase search that finds files matching a natural-language request through repeated Jev-guided passes over a directory tree.
A VS Code extension that finds past Claude Code, Codex and OpenCode conversations from a plain-language description, using a local Laya model to rerank the keyword search results.
Semantic code search for coding agents: it scans an authorized repository, asks Jev to score all eligible fragments against a plain-English question, and returns the original source excerpts with their paths and line numbers.
An agent plugin and MCP tool that sends a query plus a batch of file paths, URLs or text straight to Jev and returns relevance probabilities, so the calling agent only opens what is worth a closer look.
Uses Jev as a calibrated reranker: one relevance question per document, up to 30 documents per call, returning a probability usable both as a ranking score and an absolute threshold. Apache-2.0; benchmarked on SciFact, NFCorpus and FiQA. The repo now redirects to hev/reranker.
Command line search that asks a yes/no question of every function in a codebase and ranks the answers by the probability that the answer is yes. It is grep whose pattern is a question, not embedding search, and the author measured 1,302 functions of gin-gonic/gin in 3.7 seconds for $0.018.
Benchmark comparing Jev against Cohere Rerank 4, ZeroEntropy zerank-2 and a chat-model baseline across 14 datasets, publishing every raw API response and bootstrap ranges on each gap.
Graded relevance eval of a Jev score rerank against BM25, bge-m3 and other rankers over the Agent Skills Hub catalog, with 164 queries, 9,831 labelled pairs and the judge-circularity bias measured. Jev alone does not beat a good embedding ranker, but fusing the two wins even with Jev removed from the judging.
Lets a person or coding agent ask a plain-English question across a repository. Hunch sends each in-scope chunk to Jev and returns scored source locations, without needing to guess which words the code uses.
Measures how Laya's decision head stops attending to input past roughly 200 tokens, then fixes it by chunking documents into passages and screening each one. Tested on 400 documents, and against Jev on a public needle benchmark.
Semantic code search for coding agents built on Jev, packaged as a CLI and MCP server, with results reported on SWE-bench Lite.
A fork of the Needle Chrome extension that finds webpage passages by meaning, swapping its hosted Jev calls for Laya running locally so no API key is needed.
A natural language interface for maps: type a query like a coffee shop within 5 minutes of the aquarium and Jev scores nearby OpenStreetMap places by probability. The interactive demo runs on prerecorded responses; the source is open to run locally with live data.
Measures whether Jev reranks two Finnish legal corpora better than Voyage, GPT-based rerankers and open-weights Laya over identical shortlists and questions. Batched Jev reached 97 percent recall at 1, matching the shortlist ceiling on the MuPLeR-fi corpus.
Searches an Obsidian vault locally and offline with a BM25 engine, then reranks the top results with Jev only after you approve exactly what gets sent to it.
A video demo of intent-based email search in Gmail built with Jev, with the author suggesting a large inbox would first narrow results with semantic search or embeddings before Jev ranks them.
A single prompt pulls about 100 reference images from Cosmos, NASA and The Met using Jev, shown as a video demo of the tool.
A 27,000-parameter model that parses queries like open bugs from sam since last week into typed filters in the browser in 0.25 milliseconds. When it is unsure about a single word, it asks Jev about just that word.
Uses Jev to search @wip todo updates for specific events, such as a revenue increase or a switch to a different SaaS provider, catching entries a plain keyword search would miss.
Filters retrieved passages for relevance before they reach the generation step of a RAG pipeline.
Runs a semantic grep over a file by asking a noul question per line, turning meaning-based lookup into a batch of typed decisions.
Uses Jev scores to re-rank candidate results returned by a first-stage retriever.
Nothing matches those filters
Clear a filter to widen the search, or send us the thing you were looking for.
The usual setup has two stages. A cheap first stage, such as keyword search or embeddings, finds a shortlist. The decision model then scores each item against the question. jh1373/jev-search runs a BM25 keyword engine locally and reranks the top results only after you approve what gets sent. hev's reranker asks one relevance question per document, up to 30 documents per call.
Score gives a ranking and noul gives a keep or drop. Because the output is a probability, you can set an absolute cutoff and not only a sort order. Some entries skip the first stage. sufianetaouil/every asks a yes or no question of every function in a codebase and ranks the answers, which is grep with a question instead of a pattern.
Input length is the known limit. myxamediyar's chunklaya measures how Laya's decision head stops attending to input past roughly 200 tokens, and fixes it by screening passages one at a time. Plan to chunk long documents, and test the reranker on your own corpus before you swap it in. Another option is to rerank only a small shortlist, since each item costs a relevance question.
Yes. Entries such as hev's reranker and the TypeSafe AI re-ranking cookbook ask one relevance question per result and sort by the returned probability. Run a first-stage search to get the shortlist, then rerank it.
Ask a yes or no relevance question for each retrieved passage and drop the ones below a threshold before they reach the generating model. The TypeSafe AI cookbook on classifying RAG passages shows the pattern.
Yes. nassim-arifette/jevgrep, Bentlybro/siftr and Kelbie/hunch score code fragments against a plain-English question and return the best locations.
