Gemini

fromGemini wraps the raw Gemini client, not a pre-configured model instance, since the adapter passes model through on every call rather than baking it in up front.

GeminiClient is the public client type exported by VernLLM and accepted by fromGemini. It allows TypeScript users to type their Gemini client without depending on adapter internals. It covers both the callable model methods directly and the whole top-level SDK client (via an optional models field), so both of the forms below type-check with no cast.

New SDK (@google/genai)

The current SDK's ai.models.generateContent({ model, ... }) takes model per call, which lines up directly with how fromGemini forwards it. Pass the whole client straight through, fromGemini unwraps .models internally:

import { GoogleGenAI } from '@google/genai';
import { VernLLM, fromGemini } from 'vern-llm';

const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });

const llm = new VernLLM({
  client: fromGemini(ai),
  model: 'gemini-3.1-flash-lite',
});

fromGemini(ai.models) still works exactly as before, if you'd rather pass the narrower client:

const llm = new VernLLM({
  client: fromGemini(ai.models),
  model: 'gemini-3.1-flash-lite',
});

Tool calling

tools and tool_calls map onto Gemini's native functionCall / functionResponse parts, and toolChoice maps onto toolConfig.functionCallingConfig. Gemini 3 and later provide a native id for each functionCall; on earlier models, fromGemini synthesizes a unique id for repeated calls to the same tool. Function responses use the original function name associated with each tool call. Consecutive function response messages from a multi-tool turn are merged back into a single user content entry. See Tool Calling for the full reference.

Streaming

fromGemini implements createStream on top of the SDK's own streaming method. Text and function call parts arriving on each streamed candidate map onto text-delta and tool_call_delta chunks respectively, no extra setup is needed beyond the client already passed to fromGemini:

const { chunks, finalResult } = await llm.call({
  userContent: 'Write a short poem about the ocean.',
  stream: true,
});

for await (const chunk of chunks) {
  if (chunk.type === 'text-delta') process.stdout.write(chunk.delta);
}

See Streaming for the full stream: true contract.

Multimodal input

fromGemini supports VernLLM's provider-independent ContentBlock[] format for image input. Image blocks are translated into Gemini's native inlineData parts automatically. The selected Gemini model must support image input for multimodal requests to succeed.

const result = await llm.call({
  userContent: [
    {
      type: 'text',
      text: 'Describe this image.',
    },
    {
      type: 'image',
      mimeType: 'image/png',
      data: imageBase64,
    },
  ],
});

Image data must be base64-encoded bytes without a data: URL prefix. Supported image types are image/png, image/jpeg, image/gif, and image/webp.

Plain string userContent remains unchanged for text-only requests.