Why not call the API once per question?

A System One request carries one state and a map of questions answered in parallel, so five questions about a row cost the same single request as one. Extra questions add almost no latency or cost versus one call per question.

How can I develop without spending tokens?

Use DESCRIBE around the call. It binds the call, validates questions, and computes result columns without sending a request or costing a token.

What if different rows need different questions?

Use ask for a fixed question set with typed columns, and ask_dynamic when questions differ per row. Dynamic answers come back as JSON, not typed columns, which is what lets the questions vary.

How should confidence drive production routing?

Judge once, then split on confidence: auto-route rows at or above 0.8 and queue below for review. Confidence reflects how concentrated the distribution was, which tells you which rows you can act on without a human.

How do I install and attach it in DuckDB?

INSTALL vgi FROM community, then LOAD vgi, then ATTACH 'typesafe' with the worker launched through uvx and a CREATE SECRET of type typesafe holding your API key. Nothing is cloned.

What kinds of state can I pass to a question function?

A VARCHAR string, a STRUCT, the row alias itself, a LIST or MAP, or a JSON column with parse_json set true. Bare INTEGER or BLOB values are rejected at bind with a suggestion, not sent.

Can I judge several columns together instead of one?

Yes. Pass the row alias, such as ask(t, ...), and the model receives a JSON object of every column, so it sees the field names and can weigh a tier against a message.

CommunityJev Integrations & Adapters

Vgi TypeSafe

DuckDB table functions over a VGI worker that expose Jev Choice, Noul, Score, and multi-question ask for LATERAL joins with confidence-based routing.

See highlights

Open github.com

Stars: 4
Language: Python
License: MIT
Published by: Community
Author: Query-farm
Last verified: Oct 2026

Query-farm/vgi-typesafe

Organization repository on GitHub

View on GitHub

A VGI worker exposing TypeSafe System One questions (choice, noul, score) to DuckDB/SQL as LATERAL-joinable table functions

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

TypeSafe System One questions — choice , noul and score — as DuckDB table functions you can LATERAL join against. A VGI worker, built by 🚜 Query.Farm. Nothing to clone — uvx fetches and runs the worker, and DuckDB speaks to it over the VGI extension. Ask one question about one value. The answer comes back as columns, not prose: confidence is how concentrated the model's probability distribution was. It is the thing that makes this usable in production: it tells you which rows you can act on without a human.

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

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: Python

Highlights

One request per row issued concurrently with repeated values asked once; multi-question ask costs one request.
Confidence split routes high-confidence rows automatically and queues the rest for human review.
Pass a column, STRUCT, or whole row so field names show how tier and message relate.
DESCRIBE validates questions and result columns without sending a request or spending tokens.

Watch out

Installs via INSTALL vgi FROM community, LOAD vgi, and ATTACH with uvx; keys use CREATE SECRET and JSON columns need parse_json true to parse. Bare INTEGER or BLOB inputs are rejected at bind, and ask_dynamic returns JSON rather than typed columns. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

loki

Agent runtime with a terminal UI and a messaging gateway that keeps memory, self-built skills, session search, scheduling, and subagents across Telegram, Discord, Slack, WhatsApp, Signal, and CLI.

Jev Integrations & Adapters#community#jevbest#catalog-import

Community

wundercorp/lokiPython

Prompture

Python library for schema-enforced JSON extraction into Pydantic models, with provider-native JSON mode fallback, auto-repair, token and cost tracking, plus agents, RAG, and eval helpers.

Jev Integrations & Adapters#community#jevbest#catalog-import

Community

jhd3197/PrompturePython

hunch

Python library that maps Jev's Choice, Score and Noul onto verbs like classify, score, check, where, extract, verify and rank, keeping data shape in and out, batching one request per distinct row and caching answers.

Jev Integrations & Adapters#community#jevbest#catalog-import

Community

steven-shoemaker/hunchPython

milvus-model

The model-integration layer for the Milvus vector database, shipping adapters for embedding and reranker models from hosted providers, Jev among them, plus local backends.

Jev Integrations & Adapters#community#jevbest

Community

milvus-io/milvus-modelPython

jev-tree

Typed Python planner runtime that turns batched Jev action distributions into executable probability trees and graphs for finite decision tasks.

Jev Integrations & Adapters#community#jevbest#catalog-import

Community

Chuf-H/jev-treePython

laya-jev-GraphRAG

GraphRAG layer that runs a 4-phase ingest-to-answer pipeline over Neo4j, Memgraph, AGE or Kuzu using Laya or Jev decisions instead of per-hop LLM calls.

Jev Integrations & Adapters#community#jevbest#catalog-import

Community

bodepudimuneendra-netizen/laya-jev-GraphRAGPython

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai