Google Gemini (AI Studio)

The fastest path to Gemini — API key only, no GCP project required. Use this for quick testing or when you don't have GCP set up.

For enterprise / regulated workloads, prefer Vertex AI (data stays in your GCP project, IAM-governed).

Configuration

~/.insightworker/.env:

LLM_PROVIDER=gemini
GEMINI_API_KEY=AIza...
GEMINI_MODEL=gemini-2.0-flash-001

Get an API key

Visit aistudio.google.com/apikey
Sign in with your Google account
Create API key — pick or create a project
Copy the key (starts with AIza)

Recommended models

Use case	Model
General apps, tool use	`gemini-2.0-flash-001`
Long context, hard reasoning	`gemini-2.5-pro` (2M context!)
Cheap classification	`gemini-2.0-flash-lite`

Gemini 2.5 Pro has a 2 million token context window — useful for apps that need to load lots of documents at once. InsightWorker's per-model output cap scales accordingly (~800k chars).

Where data goes

AI Studio is a consumer-tier endpoint. Prompts go to Google's cloud, governed by Google AI Studio's data terms. For regulated workloads, use Vertex AI instead (same models, in your GCP project).

Tool calling and schema

Gemini's function-calling schema is OpenAPI 3.0-ish — stricter than full JSON Schema. InsightWorker sanitizes tool schemas automatically (strips $ref, oneOf, additionalProperties and other features Gemini doesn't support).

If a custom tool fails on Gemini but works on OpenAI, the issue is likely an unsupported schema feature. Simplify the input_schema (avoid $ref and oneOf) and re-test.

Common gotchas

Symptom	Cause	Fix
400 on tool call	Schema feature not supported by Gemini	Simplify your tool's `input_schema`
429 rate-limited	Free tier quota	Add a paid tier or switch to Vertex
Empty content with tool calls	Gemini sometimes returns tool-only responses	Agent handles this; if visible to you, it's a bug — file an issue