AI SDK

The AI SDK (packages/ai) is request-first and provider-agnostic. It does not model Conversations or persist telemetry.

Public API

• createAiSdk(initialProviders?)
• registerProvider(provider)
• complete(messages, options)
• stream(messages, options)

Core types include:

• ChatMessage: { role: "system" | "user" | "assistant", content: string }
• ProviderAdapter: requires name, defaultModel, and complete(...); stream(...) is optional
• StreamEvent canonical types:
  • response_start
  • first_token
  • text_delta
  • usage
  • request_end

Stream normalization behavior

aiSdk.stream(...) guarantees canonical lifecycle behavior even when providers are incomplete:

• If a provider lacks stream(), SDK falls back to complete() and emits: response_start -> first_token? -> text_delta? -> usage({}) -> request_end
• If provider streaming omits canonical events, SDK backfills missing response_start, usage, and request_end.
• Duplicate first_token, usage, and request_end are deduplicated.

This behavior is verified in packages/ai/src/sdk.test.ts.

Raw payload hooks

CompleteOptions and stream options support telemetry hooks:

• onRawRequest(payload)
• onRawResponse(payload)

The OpenAI-compatible adapter uses these hooks to forward provider request/chunk payloads to the caller.

OpenAI-compatible adapter

createOpenAICompatibleAdapter(...):

• uses OpenAI chat completions API
• normalizes base URL to include /v1 unless already versioned
• maps maxTokens to max_tokens
• supports buffered (complete) and streaming (stream) inference