Contextgem

ContextGem's core bet is that document extraction is a software problem, not an operations problem — and it makes that bet cleanly. The two primitives map well to how analysts actually talk about documents: aspects for the sections you care about (clauses, chapters, terms) and typed concepts for the fields you want back. The type list is deliberately narrow and useful — StringConcept, BooleanConcept, NumericalConcept, DateConcept, RatingConcept, JsonObjectConcept, LabelConcept — which means the output shape is declared rather than parsed out of free text afterwards. That alone removes a class of bugs from most in-house LLM extraction scripts. The 1.0 decision-model addition is the most interesting change. Boolean, label and rating concepts and aspects can be extracted without text generation, returning probabilities per answer. For anyone running classification-style extraction over thousands of contracts, that's the difference between a token-heavy generation job and a much smaller request. The cost guide also steers you to smaller/distilled models for basics like titles, payment amounts and dates, and warns against enabling justifications on simple extractions because they inflate generated tokens. That's the kind of advice that usually only shows up in a blog post after a customer burns money. Operationally, you get logging and configuration controls, serialization for storing and transferring ContextGem objects and results, and per-job LLM configuration so a cheap model can handle routine fields while something stronger handles the hard ones. The optimization guide covers model selection, accuracy, speed, cost and long-document handling, and there's explicit troubleshooting for small models — an admission that small models fail in specific ways rather than a marketing claim that they don't. The honest limits: there is no UI and no hosted service, so non-technical stakeholders can't touch it and you own deployment plus LLM credentials. Throughput and cost are downstream of whichever providers you configure. Long documents need deliberate tuning. And because it's a framework, the integration surface is your Python code — not a marketplace of connectors. Teams that want a dashboard and a support contract should look elsewhere; teams that want a library they can version, test and self-host will find this a reasonable foundation.