Capabilities ​
@anvia/openai implements six Anvia model and provider capabilities.
| Capability | Implementation | Notes |
|---|---|---|
| Completion | Responses and Chat Completions adapters | Both stream and support tools; Responses is the native default. |
| Embeddings | OpenAIEmbeddingModel | Batches inputs and restores provider results to input order. |
| Images | OpenAIImageGenerationModel | Requires base64 image output; URL-only responses are rejected. |
| Speech | OpenAIAudioGenerationModel | Returns bytes and a media type inferred from response format. |
| Transcription | OpenAITranscriptionModel | Uploads request bytes with the supplied filename. |
| Model listing | OpenAIClient.listModels() | Normalizes provider model metadata. |
Completion behavior ​
Both completion adapters normalize Anvia messages, documents, tools, structured output, reasoning, usage, and stream events. They expose model-aware context information for known IDs.
The adapters do not make every model support every request. Image input, reasoning controls, JSON schemas, and tool behavior remain model- and endpoint-dependent.
Responses ​
OpenAIResponsesCompletionModel is selected when no baseUrl or completionApi override is supplied. It uses provider-native Responses events and reports provider: 'openai'.
Chat Completions ​
OpenAIChatCompletionModel is selected by completionApi: 'chat', or by default when baseUrl is supplied. It reports provider: 'openai-chat'. The adapter preserves reasoning history used by compatible providers that return fields such as reasoning_content.
Embedding guarantees ​
The embedding adapter defaults to batches of 1,024. It validates integer indices, rejects duplicates and out-of-range indices, validates numeric vectors, and returns one Embedding per input in original order.
Changing dimensions changes the stored vector shape. Rebuild an existing vector index before querying it with a different dimension.
Media boundaries ​
- Image generation reads base64 output. It deliberately rejects URL-only OpenAI image responses.
- Speech returns the complete generated audio; it is not realtime voice or streaming audio.
- Transcription is a batch upload and returns normalized text; it is not a realtime transcription session.
- Video, image editing, realtime voice, batches, files, and fine-tuning are not Anvia model factories in this package. Use the underlying OpenAI client directly for native APIs outside the adapter surface.
Failure behavior ​
Provider SDK errors propagate. Adapter validation also rejects malformed embedding indices, missing image data, and transcription responses without text. listModels() wraps listing failures in ModelListingError and preserves the provider and status code when available.