Multi-model Support — References

Canonical sources for the API shapes, tokenizers, and streaming formats an abstraction layer has to reconcile.

Provider API contracts

• OpenAI API reference — chat completions, responses, tool calling, and streaming event shapes.
• Anthropic Messages API — message roles, content blocks, tool use, and stop reasons.
• Google Gemini API documentation — generate-content requests, function declarations, and safety settings.

Wire formats and schemas

• Server-sent events, WHATWG HTML Living Standard — the text/event-stream format that most streaming inference APIs are built on.
• JSON Schema specification — the vocabulary that tool parameter declarations and structured-output constraints are expressed in.
• OpenAPI Specification — useful when the internal envelope is published as a service contract of its own.

Tokenization

• tiktoken — the byte-pair encodings used by OpenAI models, and a concrete demonstration that token counts are encoding-specific.
• Hugging Face Tokenizers — tokenizer implementations and tokenizer.json artefacts for open-weight model families.