The official documentation page of TypeSafe AI, introducing how to use its model Jev and related rules.
Purpose: This documentation is intended for developers who need to use the TypeSafe AI model Jev, explaining how to call it, pricing, rate limiting, and input handling.
Main capabilities:
- Pricing method: charged by input tokens, output tokens are free. Btok is one billion tokens, Mtok is one million tokens.
- Rate limits: measured in tokens per second and requests per second; exceeding either limit returns 429 Too Many Requests. The official client SDK uses backoff retries by default and honors the retry-after header when the response includes it. If calling the HTTP API directly, you need to handle it yourself by referring to "Handling rate limits".
- Context length: Jev reads in the state at once and evaluates each question in parallel against that state. The 64k budget covers the state plus all questions combined; the 32k budget covers the state plus the single longest question. The documentation also mentions Speculative fan-out for packing multiple questions into a single request, as well as the Jev 1.13 jaggedness note on how accuracy changes as the state grows.
- Input types: Jev evaluates natural language text. Non-text inputs such as images, audio, video, and binary must first be preprocessed into text or structured fields, then sent as state; see the State documentation for supported state shapes.
- Customization methods: you can put proprietary content, records, and reference materials into the state field; encode domain rules and edge cases in the instructions and criteria within questions; break broad judgments into atomic questions, then combine the outputs in code. The documentation also mentions Composite scoring and the AutoResearch cookbook for training downstream classical models on Jev's probabilistic outputs.
- The documentation also includes sections on language support, data handling, and the model list.
Usage: provide the content to be evaluated via the state field, and submit the matters to be judged in the form of questions; you can use the official client SDK or call the HTTP API directly. The documentation provides /llms.txt as the entry point for obtaining the complete documentation index.
Applicable scenarios: evaluation, judgment, or scoring tasks that take natural language text as input and answer multiple questions in parallel against a given state;scenarios that require encoding domain rules and edge cases into questions and combining outputs at the code level.
Limitations: non-text inputs must be converted in advance into text or structured fields; rate limits are constrained separately by tokens per second and requests per second, and exceeding them results in a 429; there are two context budget specifications, 64k and 32k; the specific details of pricing, the model list, language support, and data handling are subject to the corresponding sections of the documentation.
