Capabilities ​
@anvia/gemini supports both Gemini Developer API and Vertex AI through @google/genai.
| Capability | Factory | Notes |
|---|---|---|
| Completion | completionModel() | Streaming, tools, tool choice, image and document input, structured output, and reasoning. |
| Embeddings | embeddingModel() | Task type, title, dimensions, and batching. |
| Native Gemini images | imageGenerationModel({ api: 'generateContent', modelId }) | Uses generateContent and inline image parts. |
| Imagen images | imageGenerationModel({ api: 'generateImages', modelId }) | Uses generateImages. |
| Transcription | transcriptionModel() | Sends inline audio through generateContent. |
| Model listing | listModels() | Normalizes Gemini model inventory. |
Completion ​
The completion adapter maps Anvia messages, documents, tools, tool results, structured output, usage, reasoning, and streaming events to Google GenAI content. Known model IDs expose context metadata.
Capability remains model-dependent. A completion model name can identify an image-capable or TTS-preview model without automatically making it suitable for every completion request.
Embeddings ​
The adapter defaults to batches of 100 and validates that the provider returns one numeric vector for each input. Supported task types include retrieval query/document, semantic similarity, classification, clustering, question answering, fact verification, and code retrieval query.
title and task type are forwarded to the provider. Their validity depends on the selected embedding model and task.
Images and transcription ​
The image factory selects separate provider implementations through its api discriminant. Both return decoded bytes and reject responses without valid image content.
Transcription is prompt-based audio understanding using a Gemini completion model. It maps common filename extensions to audio MIME types and rejects a response without text. It is not a realtime transcription API.
Unsupported surfaces ​
There is no Anvia speech-generation model in this package. Live sessions, files, caches, batches, tuning, and other native Google GenAI modules are not exposed as Anvia factories. Use the injected GoogleGenAI client directly when native APIs are needed.