spinf documentation

Beta. The public API is in beta; field names may still change.

spinf (speedy inference) answers many questions about a piece of content in a single pass. A question is a template that ends where the answer goes, with the answers to score there; spinf returns the probability of each answer for the content. The content is read once and every question is answered from that read, so adding questions costs only the tokens of the questions and answers themselves.

Pages

• Quickstart — get a key and make your first call
• Scoring API — templates, answers, the request and response reference
• Writing questions — content labels, answer lists, a few labelled examples, iterating on your data
• Calibration — how to use p, residual and floor_p
• Prompt-packs — ready-made question sets, like content moderation, in the same call as yours
• Limits & pricing — token limits, concurrency and billing
• FAQ — support, beta status, which models, benchmarks, determinism, media, private deployments, custom models

Concepts and jobs, step by step, are in the guides: probabilities from a language model, probabilistic if statements, zero-shot and few-shot classification and more.

Want to try before writing code? The playground runs the same calls in your browser, with ready-made examples and prompt-packs, and shows the request as code.

For coding agents

Every page is plain Markdown at https://spinf.com/docs/<page>.md, and https://spinf.com/llms.txt lists them all.