Google Gemini

Google’s Gemini family. Particularly strong on grounded reasoning and long context.

Setting the key

Gemini is one of the backends for Lettuce’s LLM-powered features (enhance, the answer RAG tool, the semantic lint advisor). It does not provide embeddings — semantic_search and embedding-based retrieval require a separate OpenAI-compatible embeddings endpoint. Select Gemini and store its key either in the GUI under Settings → AI providers, or via environment variables for the CLI.

Get an AI Studio key (AIza...) from Google AI Studio, then either paste it into Settings → AI providers (it is stored in the macOS Keychain), or export it for the CLI:

export LETTUCE_LLM_PROVIDER=gemini
export GEMINI_API_KEY=AIza...      # or GOOGLE_API_KEY

Lettuce talks to the public Generative Language API (generativelanguage.googleapis.com) and sends the key in the x-goog-api-key header.

Models

The default model is gemini-2.0-flash. Override it in Settings → AI providers, or with an environment variable:

export LETTUCE_ENHANCE_MODEL=gemini-2.5-pro

Structured output

Lettuce’s enhancement features ask the model for structured JSON. The GeminiClient maps each tool’s JSON Schema into Gemini’s responseSchema subset (uppercase type, no additionalProperties) and requests responseMimeType: application/json, so you get clean structured results back without any extra configuration.

Differences

• Gemini’s responses are sometimes less structured than Claude’s; the enforced responseSchema keeps them in line.

• Long context is useful for whole-folder enhancement passes in a single shot.

Cost

Pay-per-token, billed via Google Cloud / AI Studio. Lettuce does not track or display Gemini usage; check your spend in the Google Cloud / AI Studio console.

Related

• AI providers/01 Anthropic Claude

• AI providers/02 OpenAI