Anvia

Migrate from @anvia/openai ​

Applications that reached Azure through OpenAIClient with a custom baseUrl can move to AzureOpenAIClient. Existing @anvia/openai custom endpoints remain supported, but Azure Responses event normalization belongs to AzureOpenAIClient.

Swap the client ​

ts
// Before
import { OpenAIClient } from '@anvia/openai'
const client = new OpenAIClient({ apiKey, baseUrl: azureBaseUrl })

// After
import { AzureOpenAIClient } from '@anvia/azure'
const client = new AzureOpenAIClient({ apiKey, baseUrl: azureBaseUrl })

Replace the import, keep your Azure baseUrl and apiKey (or switch to endpoint and an Entra azureADTokenProvider), and keep the same model-handle calls. Use Azure credential variable names, such as AZURE_OPENAI_API_KEY, to keep the two providers separate.

What changes ​

  • Provider identity becomes azure-openai (and azure-openai-chat or azure-openai-responses in request traces) instead of openai, so telemetry separates the providers.
  • The Responses model resolves an omitted terminal function name from the earlier function-call item with the same item ID, per stream. Missing identities and conflicting names are rejected.
  • Azure Responses requests tag conversation messages with type: "message" for Foundry project endpoints, preserving function-call and tool-result item shapes.
  • SDK retries are disabled on the managed client.

OpenAI's own parser accepts an unnamed arguments-done event without synthesizing a name and validates the name on the completed function call. Azure's behavior is layered on through @anvia/openai/adapters, so request mapping, tool-result serialization, argument validation, and stream completion checks remain shared.

Build a similar provider ​

Provider packages that implement an OpenAI-shaped Responses endpoint can reuse the protocol adapter:

ts
import { OpenAIResponsesCompletionModel } from '@anvia/openai/adapters'

OpenAIResponsesCompletionModel exposes two protected extension points: normalizeStream(stream) to adapt endpoint-specific stream events before they are mapped, and requestParams(request) to adapt the outgoing request. Subclasses can also override provider and traceRequest(). The adapters subpath currently exports only this class; ordinary applications should keep using the root entry point.

Built for Anvia.