Inference Configuration — References

Sources defining what the generation parameters mean and how decoding behaves.

Parameter definitions

• OpenAI API reference — temperature, top_p, max_completion_tokens, stop, seed, response_format, and the finish reasons that distinguish a completed response from a truncated one.
• Anthropic Messages API — max_tokens, temperature, top_p, stop_sequences, and extended-thinking budget parameters.
• Google Gemini API documentation — generation config fields, candidate counts, and safety-setting interaction with output.
• Hugging Face Transformers generation documentation — GenerationConfig, the most explicit public description of how sampling parameters compose during decoding.

Decoding behaviour

• The Curious Case of Neural Text Degeneration (Holtzman et al., 2019), arXiv:1904.09751 — introduces nucleus sampling and explains what the top_p cut-off is actually doing to the distribution.

Constrained generation

• JSON Schema specification — the contract a constrained-decoding mode or a response-format setting is enforcing.
• Outlines — structured generation via grammar-constrained decoding, as a concrete implementation of schema enforcement at the sampling step.

Reproducibility guarantees for a fixed seed are not specified by any public document for hosted inference APIs; no reference is recorded for that claim because none is known to exist.