How is prompt injection handled?

Triage asks an injection Noul per chunk; in the recorded run a chunk instructing the model to change the rate limit scored 0.99 injection probability and was dropped before generation.

Can it run without API keys?

Yes for evaluation and scenarios: make scenarios, make eval and make test run offline on recorded calls or deterministic doubles; the full stack needs Jev and LLM keys.

What are the distinct refusal statuses?

Every refusal is a separate AnswerStatus: NO_DOCUMENTS, INSUFFICIENT_CONTEXT, DECLINED, UNGROUNDED, GENERATION_FAILED and VERIFIER_UNAVAILABLE. They are separate because the corpus being empty, nothing answering the question, an unsupported draft and the LLM being down are different bugs with different fixes.

What does a query actually cost in Jev requests?

22 for a query with 20 retrieved chunks and a 3-sentence draft: one Choice to route, one Score plus two Nouls per chunk, and one Noul per claim. Each stage batches its questions into one call, so verification latency does not grow with the number of claims.

Why does the Score answer need normalizing before a threshold?

Because Score returns the probability-weighted mean of level times probability, a position on 0 to len(criteria)-1 that can land between levels. Comparing it to a 0 to 1 threshold lets everything above level 0 through and silently turns triage into a no-op. Divide by len(rubric) - 1 instead, and a test guards that.

What did the golden set show, and how big is it?

62 labelled questions: 34 answerable, 22 unanswerable and 6 chitchat turns. Cribrix answered 0 of the 22 unanswerable against 4 of 22 for naive top-5 RAG, kept context precision at 1.00 and got 62 of 62 overall correct. The README is explicit that 0/22 carries a confidence interval up to 15%, so 62 questions show a difference rather than certify a rate.

How do I re-tune the thresholds without spending anything?

make eval-sweep re-scores the recorded signals across relevance times evidence thresholds with no API calls, making every decision in code from stored probabilities. On the recording, any evidence threshold at or above 0.5 refuses all 22 unanswerable questions while keeping a relevant chunk for all 34 answerable ones, across relevance thresholds from 0 to 0.9.

CommunityJev Integrations & Adapters

Cribrix

Precision-first RAG orchestrator that wires Jev as a System 1 gate around an LLM: it routes chitchat, triages retrieved chunks with relevance and answer-evidence Nouls, then verifies each claim and withholds drafts whose claims are unsupported.

See highlights

Open github.com

Stars: 1
Language: Python
License: MIT
Published by: Community
Last verified: Oct 2026

david96182/cribrix

User repository on GitHub

View on GitHub

A precision-first RAG orchestrator: filters before it generates, verifies before it answers.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A precision-first RAG orchestrator. It filters before it generates, and verifies before it answers. From Latin cribrum — a sieve. Most RAG systems are built to always answer. Retrieve the top-k chunks, stuff them into a prompt, return whatever comes out. This works right up until the corpus doesn't contain the answer — at which point the system confidently invents one, attaches a real citation, and the user has no way to tell.

Quoted from the project's README. Read the full README.

Tagscommunity rag hallucination pgvector self-hosted

Works forJev for Verification Jev for Decision gating Jev for Document processing

Used inJev for Software & DevTools

Languages: Python
Platforms: PostgreSQL
Capabilities: Open sourceSelf-hostedRequires API key

Highlights

A query with 20 chunks and a 3-sentence draft costs 22 Jev requests, because each stage batches its questions into one call
On a 62-question golden set it answered 0 of 22 unanswerable questions versus 4 of 22 for naive top-5 RAG, and kept context precision at 1.00
Every refusal is a distinct status: NO_DOCUMENTS, INSUFFICIENT_CONTEXT, DECLINED, UNGROUNDED, GENERATION_FAILED or VERIFIER_UNAVAILABLE
Choice confidence gates routing at 0.8; a real question Jev labelled CHITCHAT at 0.49 confidence was searched and answered instead
make eval replays recorded live calls byte-for-byte offline and CI gates on the numbers; make eval-fake is only a wiring test

Watch out

MIT license. The golden set is 62 questions, so the 0/22 result has a confidence interval up to 15%; the README explains where the naive baseline is given the benefit of the doubt.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

OpenViking

39.6k

Volcengine's agent context database that unifies memory, knowledge RAG and skills under a viking:// virtual filesystem; it ships a Jev rerank client that scores candidate documents against api.typesafe.ai and treats the probability as relevance.

Jev Integrations & Adapters#community#agent-memory#rag

Community

volcengine/OpenVikingPython

jev4pg

224

A Jev harness for PostgreSQL: 41 semantic operators plus a native Rust extension evaluate row-level predicates with Noul questions and keep the returned probabilities for explicit decision thresholds, leaving joins and aggregation to SQL.

Jev Integrations & Adapters#community#postgresql#nl2sql

Community

Sheltercosmo/jev4pgPython

AutoGPT

187.5k

Autonomous agent platform with 187,504 stars for building and running agents from plain English or a visual builder. Its platform features first-class TypeSafe Jev decision blocks for typed routing, filtering, scoring and confidence-gated next-action dispatching.

Jev Integrations & Adapters#community#autonomous-agents#platform

Community

Significant-Gravitas/AutoGPTPython

Inbox Zero

12.4k

Open-source AI email assistant (Next.js, Postgres, 12k+ stars) with a decision-model layer that can run on openrouter:typesafe/jev-1.13: Jev-backed questions cover rule selection, cold-email detection, thread status, recurring patterns, unsubscribe pages, and sender categorization, each with evals.

Jev Integrations & Adapters#community#email#gmail

Community

elie222/inbox-zeroTypeScript

GPTCache

8.2k

A semantic cache for LLM queries, fully integrated with LangChain and llama_index. Zilliz's cache integrates TypeSafe Jev Noul checks to evaluate cache hit freshness and time-dependent query validity, cutting repeated requests and API cost.

Jev Integrations & Adapters#community#semantic-cache#langchain

Community

zilliztech/GPTCachePython

Vector Graph RAG

255

Graph RAG framework over Milvus vector search with no graph database required: Jev Noul judgments select graph relations and rerank full source passages alongside vector-search candidates, reaching 86.07% average Recall@5 on 1,000 MuSiQue and 1,000 2Wiki queries in the two-stage evaluation.

Jev Integrations & Adapters#community#rag#retrieval

Community

zilliztech/vector-graph-ragPython

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai