User repository on GitHub
Jev adapter for Mellea
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Integrations & Adapters
Jev4Mellea
Python adapter that puts Jev's Noul, Choice and Score checks inside Mellea generation flows: verify a requirement, classify text into caller-defined categories, or rate it on an ordered scale, each attachable as a Mellea Requirement.
- Stars
- 3
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- SoundBlaster
- Last verified
- Oct 2026
About this project
A small Python adapter that brings TypeSafe Jev's semantic checks into Mellea. Use Jev to verify generated text, classify it into your labels, or rate it on a scale. Mellea continues to manage generation and repair. Verify requirements with Jev Noul and configurable accept/reject thresholds. Classify text with TypeSafe Choice and caller-defined categories. Score text on an ordered scale with TypeSafe Score. Ask several questions in one request and inspect returned token usage.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Each check converts into a Mellea Requirement, so a Jev judgment can gate a generation and repair loop instead of only annotating finished text.
- Two configurable Noul thresholds bracket an uncertain band that is never silently accepted; outcome is pass, fail or uncertain.
- An optional LayaProvider runs the same checks against a local Laya-MLX checkpoint on Apple Silicon, mapping its answers into the same contracts.
- make demo runs against a mocked Jev response, so the example works with no key and no network access.
Watch out
It is an unofficial, synchronous adapter, and the README is careful that Jev evaluates text while Mellea still owns generation and repair. The Makefile defaults to Python 3.13 while the project targets Python 3.11+, so another interpreter needs make install PYTHON=python3.11 or setting PYTHON; the badges pin Mellea 0.7.0 and TypeSafe SDK 0.7.x. On accuracy the adapter's own wording is that threshold values are application policy, not accuracy guarantees, and that Noul returns no separate confidence field and no textual explanation to quote. LayaProvider is a different model with different calibration: Laya derives Choice confidence from normalized entropy, so minimum_confidence has to be recalibrated per backend rather than copied from Jev, and the checkpoint's license is separate from the Apache-2.0 Laya-MLX runtime. Score accepts 2-10 ordered descriptions, Choice up to 255 labels, and reported token usage is informational only. The README continues past the point this listing's source text ends, cut off mid-section at examples/evaluate.py, so limits stated after that are not reflected here. 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.
Just a check. The README states it is an unofficial synchronous adapter where Jev evaluates text and does not generate or repair it, while Mellea continues to manage generation and repair. What the adapter adds is the ability to hand a Jev check to Mellea as a Requirement so it can take part in sampling: pass the requirement into instruct() alongside a strategy such as RepairTemplateStrategy, then read the result through accepted_text() rather than sampled.result, which can come from a failed or incomplete sampling.
You get uncertain, and it is not silently accepted. JevVerifier takes two configurable thresholds, accept_at and reject_at, and the README says the space between them is uncertain and is never silently accepted, with outcome one of pass, fail or uncertain. The worked example routes uncertain to another check or human review. The adapter is explicit that those thresholds are application policy, not accuracy guarantees.
Yes, through the optional Laya backend. On Apple Silicon macOS, pip install -e '.[laya]' and laya_mlx.load a checkpoint, and LayaProvider wraps an already-loaded agent so importing the base package does not pull in Laya or MLX. Its Noul, Choice and Score answers are mapped into the same response contracts JevClient returns. The catch is calibration: Laya computes Choice confidence from normalized entropy, so recalibrate minimum_confidence for that backend rather than reusing a Jev threshold.
None for the demo. Once project dependencies are installed, make demo runs against a mocked Jev response with no key and no network access. Live requests are separate and the README warns they may incur charges, with the key coming from the TypeSafe console as TYPESAFE_API_KEY. One practical trap in the generation example: keep the JevClient open until sampling finishes, because the requirement calls it during validation.
Verify a requirement with Noul and configurable accept or reject thresholds, classify text with Choice into caller-defined categories, or score it on an ordered Score scale. Requests can ask several questions at once and inspect token usage, and any check can be attached to a Mellea Requirement so it participates in sampling. Mellea keeps managing generation and repair.
evaluate returns an outcome of pass, fail or uncertain plus p_yes. Two configurable thresholds define the uncertain band in between, which is never silently accepted as a pass. Noul does not return a separate confidence field or a textual explanation.
Yes. Install the optional provider with pip install -e '.[openai]' and set OPENAI_API_KEY. The OpenAI Decisions API provider is in public beta, supports Noul verification, Choice classification and Score rubrics, but only one question per request.
More Jev resources like this
25GitHub stars
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
wundercorp/lokiPython51GitHub stars
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
23GitHub stars
Memory-maintenance library that checks every stored memory against each new event with Jev yes-or-no votes, then marks stale facts superseded or contradicted instead of editing them.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
chopratejas/invalidatePython
chopratejas/invalidatePython5GitHub stars
Community Agent Zero plugin that exposes one typesafe_query tool for combined Choice, Noul, and Score questions over the same evidence, with probability cards in chat.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
3clyp50/a0-typesafe-aiPython
3clyp50/a0-typesafe-aiPython4GitHub stars
Four small Pydantic AI demos that put a Jev question at a fixed hook, before the model runs, before a tool runs or after the run ends, and act on the number.
Jev Integrations & Adapters#community#jevbest#catalog-import
1GitHub stars
A Letta agent skill that wraps TypeSafe System One (Jev) as a cheap external judge for texts and decisions, plus a paragraph scanner that maps where a document argues, lists, asks, or admits uncertainty. The README is in Russian.
Jev Integrations & Adapters#community#jevbest#catalog-import
From the Jev community
Posts from builders shipping with Jev right now.