StackMap
Subscribe

contextgem vs kev

Declarative LLM extraction from documents: describe Aspects and Concepts in plain language, get structured values back with paragraph- or sentence-level references and built-in justifications. — versus — Open reproduction of TypeSafe's Jev: a LoRA + readout head on Qwen that answers many typed questions about one document in a single prefill pass, returning calibrated probabilities.

The curated verdict

Both turn one document into structured typed values; contextgem prompts an LLM declaratively per Aspect/Concept, kev trains a head that answers every question in one pass and returns calibrated probabilities instead of generated text.

contextgemkev
Stars2.0k492
Forks18630
LanguagePythonPython
LicenseApache-2.0Apache-2.0
Last activity1 months agotoday
Topicsocr, datatraining, local
Curated connections108

contextgem — the curator's take

Its thesis is explicitly anti-RAG: stop chunking and retrieving, put the whole document in a long context window and extract in place. You declare Aspects — the sections or themes worth pulling — and Concepts — the entities, booleans, numbers, ratings and JSON objects inside them — and the framework writes the prompts, builds the validation models, and returns every item with reference_paragraphs, reference_sentences and a justification. That provenance is the point: it makes contract and filing review defensible rather than merely plausible. Two limits it states about itself: no cross-document querying or corpus-wide retrieval, so reach for a real RAG framework there; and its detailed instruction style wants a model at gpt-4o-mini level or better, so 8B local models will struggle even though LiteLLM happily points it at Ollama.

kev — the curator's take

Reach for kev when the task is decisions, not prose: route this ticket, score this risk, answer 30 yes/no questions about one document — and you want a number you can threshold on rather than text you have to parse. The head is trained with cross-entropy on labelled outcomes, so the probabilities mean something (kev-8b: Brier 0.34, 8% confident errors out of domain), and a block-causal mask keeps questions from seeing each other, verified to 4e-6 against separate requests. It speaks TypeSafe's System One API, so their SDK works against localhost with a base_url change. Not a general chat or agent model - it never decodes, and it only answers the question types you define (noul/choice/score). Don't use it zero-shot on your own domain either: the value is in training the head on your labels, and the three 0.6B/4B/8B checkpoints are explicitly previews that failed the author's own release screen on held-out rule reasoning. Real Jev still wins out of domain (0.86 vs 0.77) - kev's pitch is that it runs on your laptop and you own the weights.