Jev for search and RAG
Jev does not search. It judges what something else already found: where a query should go, which result goes first, which passages the writer gets to read. These are the builds that use it that way, grouped by job, with the per-query costs their authors published and one careful test where it didn't win on its own. Source: TypeSafe re-ranking cookbook.
Updated 8 Oct 2026 · by Made with Jev
In short
- Keep BM25 or embeddings for recall. Reranking reorders a shortlist and cannot add a passage fast search missed. rerank cookbook
- On TypeSafe's CLERC test, reranking 30 BM25 candidates moved the right passage to first place for 18% of queries, up from 5%, on jev-1.12. rerank cookbook
- In Jason Zhu's 164-query test, Jev alone barely beat vector search. Fused with it, it won under all three judges. x.com/GoSailGlobal
- For RAG, ask several Noul questions about each passage and decide in code. TypeSafe's own cookbook says its injection check is "a filter" and that "nothing here is a security boundary". classifying_rag_passages cookbook
What one search costs
| Source | What was ranked | Published cost |
|---|---|---|
| TypeSafe re-ranking cookbook | 30 BM25 candidates for each of 40 queries, 1,200 calls, $0.0645 in total | $0.0016 per query |
| YC Indexor (Aayan) | Up to 320 finalists scored in one request | Roughly one second and about a fifth of a cent |
| siftr (Bently) | Large repo search | 1–2¢ per search |
| Jason Zhu, fusion test | One extra API call per query, about $0.0002. The whole evaluation (164 queries, 9,831 graded pairs) cost $2.6 | ~$0.0002 per query |
Jev bills input tokens only, at $0.042 per million as of Oct 2026, so a search costs whatever amount of candidate text you send. See Jev pricing.
Work out what the query is asking
Before anything is fetched, Jev reads the query and picks the sources, the time window or the intent. superagents-lab's Jev Search does this for web search and shows relevance scores instead of writing an answer.
Ask for the intent as a Choice, and ask everything the next step might need (a date range, a source, a language) in the same request. Code ignores the answers it doesn't use, and each extra question costs only its tokens. Sources: Ask speculative questions and intent-routing pattern.
GitHubSearch
Jev Search
Web search where a decision model picks the sources and ranks them.
superagents-lab
516
ギガビット@ゲームつくるひと
@gigabit_million
これGoogle検索でいいのではと思いながら作ってたけど、Chrome拡張機能でJevで意図で検索できるの作ってる人がかなりいたみたいでJevはGoogle検索より良いケースあるみたい。Jevのユースケースとして良いやつだった
nader dabit
@dabit3
Jev is really good at intent-based search! How it looks in Gmail: (for a huge inbox you'd prob let semantic search / embeddings pull first but still much better experience)
XSearch
Intent-based search in Gmail
GitHubRobotics and devices
home-assistant-typesafe
Home Assistant routes what you said through Jev, not an LLM.
Allen Porter
2Send the query to the right place
The same read, used to pick a destination: a query engine, a tool, or the one set of docs worth loading. Gipp's pipeline drops 14,000 of 20,000 events a day as noise and sends 500 to a person, for $0.84 of Jev.
TypeSafe's intent-routing pattern classifies the request once. One intent goes to plain code with no LLM, two go to different specialist LLMs, and a complexity Score with a confidence check decides between an LLM and a person. Source: intent-routing pattern. Jev router for model routing.
Burak Karakan
@burakkarakann
I built a query layer that automatically routes the queries to one of the configured query engines using Jev, all running on top of Iceberg. Fast decision models unlock interesting use cases indeed. Should we open-source this?
XRouting and model choice
A query router on top of Iceberg
CJ (Coding Garden)
@CodingGarden
I built a chat bot with jev, no LLM at all! Responses are instant, no hallucinations. I hooked it up to web search, wikipedia, weather, todoist and home assistant. Jev decides what tool to call and what args to use based on the prompt. Instant answers cite sources as well!
XRouting and model choice
A chat bot with no LLM
NatashaTheRobot
@NatashaTheRobot
I built a small AskHIG CLI tool + agent plugin to help the agent load only the RELEVANT HIG guides into context via an /askHIG command. It uses @typesafeai Jev to decide which guides, so all you need is a personal typesafe API key and you're all set! I would open source it, but don't want to get in trouble with 🍎... If you'd like access, reply with your github username in the comments and I'll add you to the repo!
XContext and memory
AskHIG
Coach Shweta Bajaj
@shwetabjaj
Jev Skill Suggestion for Claude Code is a smart idea. Instead of loading every skill into context, Jev decides which one is actually relevant and injects only that. My takeaway: better context hygiene, less clutter, and potentially more efficient agent workflows. @typesafeai @vercel
XContext and memory
Skill suggestion injects one skill
Gipp 🦅
@gippp69
Jev builders: the model bill is decided before the model is called /routing 20,000 events/day become: → 14,000 die as noise → 4,700 become logs → 800 earn a draft → 500 go to human review only 4% reaches writing. Jev costs $0.84; the full loop lands at $29.04. optimize the 4% before the writer.
XRouting and model choice
20,000 events a day, 4% reach the writer
- Events per day
- 20,000
- Reach the writer
- 4%
- Jev cost
- $0.84
- Full loop
- $29.04
Rerank a shortlist
Keyword or vector search builds the shortlist, and Jev scores each candidate against the query. Bently's siftr found the right file in the top five 82% of the time on SWE-bench Lite, against 52% for BM25.
The cookbook asks one Noul question per query–candidate pair and sorts by the value it returns. Over 40 CLERC legal queries, each with a 30-candidate BM25 shortlist, top-1 went from 5% to 18%, top-5 from 15% to 35% and top-10 from 38% to 62%. That was 1,200 calls on jev-1.12. The cookbook asks one question per call for clarity, and notes that a real application would ask several about the same pair in one call. Source: rerank_typesafe cookbook.
GitHubCoding and code review
siftr
Semantic code search for coding agents, benchmarked on SWE-bench Lite.
Bently
82%
Aayan
@aaayandev
Jev (@typesafeai) is so insane & cheap for search!! > 6000+ @ycombinator Startups indexed. > Sub 1 second search results. > 90M tokens & $2.7 in total testing costs. Search any startup in a second, in any way! - Color - Niche - Your Competitor - Age - Image - etc... > watch the entire video, it's so freaking cool omg! > this is the coolest thing i have ever built for fun! (worked on it for 2 days straight!)
Ian Nuttall
@iannuttall
Cloudflare Workers has Jev now so I'm putting it to the test on keep.md - 7x faster search rerank compared to the current hybrid - 50x faster tagging of content vs GLM 4.7 Flash with no failures
GitHubSearch
ensk
An English–Icelandic dictionary with Jev reranking the results.
Jökull Sólberg Auðunsson
0
GitHubSearch
jev-search
Offline Obsidian search, with an optional Jev rerank you approve first.
jh1373
1
Saksham Malhotra
@SakshamMalhot27
Launching JevRelevanceRetriever just $0.73 across 6,460 calls Open src Jev LangChain BaseRetriever for rel scoring Install and use it, no chain changes ;) pip install jev-relevance Repo-https://github.com/saksham-malhotra-27/jev-relevance Open for issues and PRs, feel free to contribute
XSearch
PickJevRelevanceRetriever
Likit D
@Likitd_
Jev jev jev jev Your RAG pipeline just got a relevance upgrade. excited to introduce jev-ranker — Jev-powered reranking and relevance filtering How much can smarter reranking improve a RAG pipeline? Try : npmjs.com/package/jev-ra… Benchmark Results:
XSearch
jev-ranker
Viv
@Vtrivedy10
Jev for RAG in almost all cases you trust the semantic matching capability of Jev more than dot product similarity very useful as the direct similarity metric in small data cases and a great reranker with big data
XSearch
Jev for RAG
Decide which passages reach the writer
Retrieval returns passages that look like the query, including some that contradict it or carry instructions. These builds ask Jev about each passage and drop or hold it back in code before the LLM reads anything.
Put the query and one passage in the state, and ask four Noul questions: is it relevant, does it contain answer evidence, does it contradict the query's premise, and does it contain a prompt injection. In code, test them in a fixed order: injection above 0.70 → drop; contradiction above 0.70 → conflict block; relevance below 0.45 → drop; evidence above 0.55 → include; anything else → drop. TypeSafe chose those numbers for its own corpus and calls them a starting point. The writer gets evidence and conflicts in separate blocks, so it can push back on a false premise. In the cookbook, cosine similarity ranked the planted injection first, at 0.584, and its injection score of 0.99 removed it. That is one request per passage, so cost grows with k. Run on jev-1.12, 2026-08-27. Source: classifying_rag_passages cookbook.
After the answer is written, a Choice per citation (verified, unsupported, contradicted or fabricated) checks it against the source. A plain string match catches quotes that aren't there at all. In the cookbook, all four planted failures were caught, and the four accurate citations came back verified at confidence 0.93 or higher. Source: citation_check cookbook.
N01ennn describes a variant that scores all 20 hybrid-search candidates in one request, with answerable and injection gates, and passes only passages above 0.6. That is N01ennn's description; we haven't found the research group's own write-up. x.com/N01ennn. See adversarial content.
NO1ennn
@N01ennn
this is pure f*cking treasure A Stanford AI research group finally drew the perfect RAG system: retrieval, Jev and agents in one loop, and it fixes the 3 things that break every RAG app: > the LLM reads 20 passages when only 3 matter > it answers questions your docs can't answer > it trusts whatever text it retrieves here's how it runs: > a lead agent sends the query > hybrid search pulls the top 20 candidates, dense + keyword > ONE Jev request scores all 20 + 2 gates: answerable? injection? > only passages above 0.6 reach the writer agent > a second Jev call checks every claim against its source > grounded answer, with citations and when the docs don't have it: > answerable fails, the writer never runs > a researcher agent rewrites the query and retries once > still nothing? "not in the docs". zero tokens spent on a guess retrieval casts the net. Jev decides what's real. agents do the work save this before you build your next RAG
GitHubContext and memory
Jev Sift
Classify first, read selectively: an agent plugin and MCP tool.
Kush Bhuwalka
46
GitHubSDKs and integrations
jev-mcp
An MCP server with claim checks, screening and ranking built on Jev.
Joey Kudish
506
GitHubContext and memory
perfectrecall
Agent memory where Jev decides what is worth recalling.
Arslan R.
2
たく｜ガチのCopilot達人
@taku_ai_case
これはめちゃくちゃ勉強になりました。 Jevに回答させるのではなく、CodexやClaude Codeへ渡す記憶候補を選ばせる。文章を書けないAIの特性を、ここまで実用的な仕組みに落とし込む発想がすごいです。
XContext and memory
Let Jev pick the memories, not the answer
GitHubContext and memory
winnow
A context sieve for Claude Code: Jev judges each tool result before it lands.
Ghaleb Dweikat
103Label results so people can filter them
Tags the source never had, added in one pass over the collection. Justine Moore classified thousands of Zillow listings by architecture, renovation and distance to a freeway in under 20 seconds, for $0.18.
Run the labels once over the collection, store them, and filter in code. Every label can be its own question in the same request. Source: primitives. See Jev for classification.
Justine Moore
@venturetwins
Jev can serve as better natural language search on websites. It can scan thousands of Zillow listings and classify properties by things you can't normally filter for - e.g. architecture, renovation status, proximity to freeways. This was done in <20 sec and costs $0.18 👇
Jon Kraayenbrink
@kraayenJon
Jev + LinkedIn bookmarks = your saved posts are finally searchable ✅ 479 posts classified in 225 seconds for $0.0195 🤯 LinkedIn has no native way to search what you saved Jev reads every post and tags the topic, hook, and format you can finally find that post you saved 6 months ago So: save, classify, search. Never lose a good post again. Free and open source. try it below ↓
Meet Shukla
@meetshukla_
i’m probably going to regret this. 10,000+ videos were imported by our users into Ghostfeed in last 6 months. this weekend, we picked 1,066 reaction videos and analysed all of them with JEV + GEMINI. now you can search by structure, reaction, and who’s on camera. opening the library today which has been our months of data gathering, clone all videos to mass publish them on instagram, tiktok and youtube. comment "reactions" if you want the access to the library.
Marc Köhlbrugge
@marckohlbrugge
Using Jev to filter through my @wip todos It allows me to super quickly find all the instances where I increased revenue, got stuck, switched to a different SaaS provider, etc Things a regular keyword search would never catch
Search by meaning without an index
Each line, row or file gets a typed question at query time, so there is no embedding store to build or refresh. uehaj's sys1grep scores every line against a meaning and combines meanings with AND, OR and NOT.
Give each line an ID and make the IDs a Choice question's options. Choice probabilities always sum to 1, so some line always ranks first. A Noul in the same request asks whether the document answers the query at all. A Choice takes up to 255 options; past that, use two passes, first picking a window and then a line within it. The cookbook searches GitHub's Terms of Service, split into 218 lines. Source: semantic_find cookbook.
GitHubCoding and code review
jevgrep
Semantic code search for agents: find behaviour, not strings.
Nassim Arifette
101
GitHubCoding and code review
blink
Codebase search by sending a hundred walkers through the file tree.
Ellipsis
99
Zachi
@iam_zachi
I think I just cooked something 🔥 jev(): a PostgreSQL extension that searches your whole database in natural language. No index, no embeddings, just one function. WHERE jev(people, 'could work from home') or WHERE jev(people, 'name sounds european') 129 rows judged in ~1s for $0.0009. Second run: 6ms from cache.
GitHubSDKs and integrations
duckdb-jev
Ask a question about every row in SQL, and get a real SQL type back.
Colliber
28
GitHubSearch
Jev-powered Obsidian search
Search an Obsidian vault with Jev deciding relevance.
MikeLembo
12
AYi
@AYi_AInotes
你是不是也在推特收藏了上千个书签，最后只存不看， 开发者 Alex 用jev解决了这个痛点， 他刚上线的这个工具 Margin, 算是真正把你的书签收藏救活了， 以往尝试用传统大模型做个人书签检索, 体验几乎是灾难级的: 把几百条推文全文丢给大模型或者向量库, 每次搜一句话要在云端等上半分钟, 还要硬烧几块钱的调用费, 根本没人愿意天天用。 而接入专做强类型决策的 Jev 之后, 整个检索逻辑彻底被掀翻了。 你只要用扩展把这几年积攒的书签一键导出导入工具, 接下来直接用最口语化的人话去搜, 比如问它有哪些关于冷启动获客的实操技巧。 系统完全不需要自回归逐字生成废话, 而是由底层的决策小脑以单条几十毫秒的速度在后台做并行打分。 短短几秒钟之内, 上百条沉睡了几年的推文被挨个计算出匹配置信度, 界面带着进度条瞬间弹出一份按相关性精准排序的推文清单, 甚至把每条内容到底契合在哪用数字标得清清楚楚。 百万 Token 只要四分钱且输出免单的机制, 让这种以往极度奢侈的遍历打分, 变成了一次只需要花几分钱的日常微小操作。 真正好用的知识库从来不是逼着人类自己去建复杂的分类文件夹, 而是把检索成本压低到可以肆无忌惮地用自然语言反复提问。 以前是把好内容随手扔进黑洞当个数字囤积狂, 从今天起是用毫秒级的决策神经随时把冷数据捞出来变成生产力。 https://x.com/alexchristou_/status/2101674202361221376/video/1
XSearch
Margin, bookmark search
Live siteSearch
A forum database, searched by Jev
Jev finds the most relevant posts in a forum database.
bigmuzzy
Where it lost: an honest negative
Jason Zhu's test, posted in Chinese. x.com/GoSailGlobal
- The test used the 33,047-entry Agent Skills Hub catalogue, 164 real queries in Chinese and English, and 9,831 graded pairs. It cost $2.6.
- Jev reranking bge-m3's top 30 added 0.012 NDCG@10, with a confidence interval that crosses zero. MRR and top-3 hits rose clearly: it is good at putting the best result first.
- Judge bias: Jev also helped grade the test set. With Jev as the only judge it led by 0.053. With both judges combined it led by 0.012. With Haiku as the only judge it trailed by 0.028. Zhu trusts only the Haiku column. Lesson: don't let the reranker grade itself.
- RRF fusion of the Jev and bge-m3 rankings reached NDCG@10 0.864. That beat pure vector search by 0.064 to 0.116 under all three judges.
- Their own keyword search scored 0.609. Half of the relevant results never entered the candidate pool, and no reranker recovers those. (The TypeSafe cookbook makes the same point.)
Jason Zhu
@GoSailGlobal
拿 Jev 做搜索重排，我先泼一盆冷水：单独用，它没打赢向量检索 TypeSafe 的 Jev 这阵子很火，一堆项目拿它做重排。我们在 Agent Skills Hub 的 33,047 条目录上认真测了一次，164 条中英文真实查询，9,831 对分级标注，整套只花了 2.6 美元 三个结论 01｜单独重排，约等于没赢 Jev 重排 bge-m3 的前 30 条，NDCG@10 只多了 0.012，置信区间跨过零。MRR 和前三命中率倒是涨得明显，它很会把最强的那一条顶到第一，后面几条基本是重新洗牌 02｜裁判偏差，被我们量出来了 Jev 自己也参与了打标注，这就是循环 只用 Jev 当裁判，它领先 0.053 两个裁判合并，领先 0.012 只用 Haiku 当裁判，反而落后 0.028 同一组比较，换个裁判结论直接翻面。所有涉及 Jev 的结论，我们只认 Haiku 那一列 03｜真正稳赢的是融合 把 Jev 和 bge-m3 的排序做 RRF 融合，NDCG@10 到 0.864，比纯向量高 0.064 到 0.116，三种裁判下都成立。代价是每次查询多一次 API 调用，约 0.0002 美元 顺便暴露了我们自己的问题：Hub 线上 CLI 用的关键词排序只有 0.609，短板是召回。相关结果有一半压根没进候选，后面怎么重排都救不回来 数据、标注、每条查询的得分全部开源，不用 API key 就能复现打分
Guidex.com
Jev as a reranker: an honest negative
Jason Zhu tested Jev reranking on 164 real queries. Alone it did not clearly beat vector search; fused with it, it did. In Chinese.
Where Jev doesn't fit in search
- It doesn't retrieve from a corpus; it ranks what you pass it. rerank cookbook
- Text only. Jev images.
- The context limit is 64k tokens per request, and 32k for the state plus the longest question, so send passages, not whole documents. docs.typesafe.ai/models. See large state.
- English is the primary training language and other languages score lower. Zhu's test was mixed Chinese and English. language support
- Rate limits are 100K tokens/s and 80 requests/s, "adjusting dynamically" (as of Oct 2026). A one-request-per-passage design uses k requests per query. docs.typesafe.ai/models
Reading
Jason Zhu
@GoSailGlobal
拿 Jev 做搜索重排，我先泼一盆冷水：单独用，它没打赢向量检索 TypeSafe 的 Jev 这阵子很火，一堆项目拿它做重排。我们在 Agent Skills Hub 的 33,047 条目录上认真测了一次，164 条中英文真实查询，9,831 对分级标注，整套只花了 2.6 美元 三个结论 01｜单独重排，约等于没赢 Jev 重排 bge-m3 的前 30 条，NDCG@10 只多了 0.012，置信区间跨过零。MRR 和前三命中率倒是涨得明显，它很会把最强的那一条顶到第一，后面几条基本是重新洗牌 02｜裁判偏差，被我们量出来了 Jev 自己也参与了打标注，这就是循环 只用 Jev 当裁判，它领先 0.053 两个裁判合并，领先 0.012 只用 Haiku 当裁判，反而落后 0.028 同一组比较，换个裁判结论直接翻面。所有涉及 Jev 的结论，我们只认 Haiku 那一列 03｜真正稳赢的是融合 把 Jev 和 bge-m3 的排序做 RRF 融合，NDCG@10 到 0.864，比纯向量高 0.064 到 0.116，三种裁判下都成立。代价是每次查询多一次 API 调用，约 0.0002 美元 顺便暴露了我们自己的问题：Hub 线上 CLI 用的关键词排序只有 0.609，短板是召回。相关结果有一半压根没进候选，后面怎么重排都救不回来 数据、标注、每条查询的得分全部开源，不用 API key 就能复现打分
Guidex.com
Jev as a reranker: an honest negative
Jason Zhu tested Jev reranking on 164 real queries. Alone it did not clearly beat vector search; fused with it, it did. In Chinese.
Guideyoutube.com/@engineerprompt
Steerable Reranking: How JEV Solves RAG
Prompt Engineering puts Jev in the reranking step of a RAG pipeline: why vector search and cosine similarity fail, how a steerable reranker works, and how it compares with LLM rerankers and cross-encoders. A Colab notebook comes with it.
Isaac Flath
@isaac_flath
I've been using Jev by @typesafeai Here's the six things i've tried and am confident I'll still use Jev for 60 days from now. There's many more experiments, ideas, and things I think I will use it for. It's a big deal (more on why in next post). But I am only sharing things that I am 99% sure will lead to stuff I will still be using Jev for in 60 days. That means I started with small, boring, but useful, stuff. - Fact-checking my scripts - Ranking my news feed - Finding the right text in PDFs - Checking citations - Grouping my review notes - Figuring out why agents fail (eval over traces) https://isaacflath.com/writing/six-things-i-tried-with-jev
Guidex.com
Six things I will still use Jev for
Isaac Flath’s shortlist: fact-checking scripts, ranking a news feed, finding text in PDFs, checking citations, grouping notes, and evals over agent traces.
Ricky Grannis-Vu
@RickyGrannisVu
Jev gives you a probability. You set two lines on it: above the top one it's a yes, below the bottom one it's a no, anything in the gap runs a third branch. It's just a normal branch, so you can handle uncertainty flexibly.
Guidex.com
Jev’s probability thresholds
Jev outputs a probability, two thresholds split the answer into yes, no and a third branch between them.
delost
@thedelost
this is pure f*cking treasure 10 GitHub projects that put Jev between your agent and every decision your agent should stop paying a frontier model to answer yes or no ROUTE THE WORK 01 jev-router > https://github.com/gargpratyush/jev-router 02 jcm-router > https://github.com/adarshmishra07/jcm-router 03 jev-agent-skill-router > https://github.com/GodsBoy/jev-agent-skill-router GUARD EVERY ACTION 04 pi-heed > https://github.com/Nyarlathoteppppp/pi-heed 05 fx > https://github.com/vercel-labs/fx 06 Foreman > https://github.com/thruwire/foreman SCORE WHAT MATTERS 07 LlamaIndex Jev > https://github.com/WiktorB2004/llama-index-jev 08 citation-verifier > https://github.com/MarissaFamularo/citation-verifier PUT IT TO WORK 09 yoshi > https://github.com/compozy/yoshi 10 mastra-jev-moderation > https://github.com/CodeAlive-AI/mastra-jev-moderation the loop: your agent hits a fork > Jev answers with a probability > your code decides what that number allows > the agent moves on 3 builds I'd explore: coding agent: jev-router > fx > Foreman > yoshi support desk: jev-agent-skill-router > mastra-jev-moderation > LlamaIndex Jev research: jcm-router > pi-heed > citation-verifier save this, then build a decision layer for your agents ⭣
Guidex.com
10 GitHub projects that put Jev between your agent and every decision
delost's list of ten repos, grouped as route the work, guard every action, score what matters and put it to work, with three suggested stacks for coding, support and research agents.
Common questions
- Can Jev replace a vector database?
- No. It scores the candidates you give it. Keyword or vector search still builds the shortlist, and reranking can't add a passage that search missed.
- Is Jev a good reranker?
- On TypeSafe's CLERC test it raised top-1 from 5% to 18% over BM25. In Jason Zhu's 164-query test it barely beat vector search alone, but fusing the two rankings beat vector search by 0.064 to 0.116 NDCG@10.
- What does reranking with Jev cost?
- TypeSafe's cookbook paid $0.0016 per query for 30 candidates. Builders report about a fifth of a cent (YC Indexor) to 1–2¢ on a large code repo (siftr).
- How do I filter RAG passages with Jev?
- Ask several Noul questions about each query–passage pair (relevance, answer evidence, contradiction, injection), then pick thresholds in code to include the passage, flag it as a conflict, or drop it.
- Does Jev stop prompt injection in retrieved text?
- No. TypeSafe calls its injection question a filter, not a security boundary, so the writer has to treat every passage as untrusted.
- Can Jev search a document without embeddings?
- Yes, for up to 255 lines per Choice question. Use a Noul in the same request to check whether the document answers the query at all.
Made with Jev is independent and not affiliated with TypeSafe AI. Every figure on this page is the one its author published, linked to where it can be checked.