Integrations / Provider Guides

Gemini

Connect Aurora to Google Gemini for Pro, Flash, and embedding models over the OpenAI-compatible endpoint.

Browse docs

Use this provider to route to Google's Gemini models. Aurora talks to Gemini through Google's OpenAI-compatible endpoint, so the wire format matches what your OpenAI SDK already expects.

Activate

text

GEMINI_API_KEY=...

Restart Aurora. The gemini provider appears at /v1/models with the current Gemini catalog.

Default models

Model	Use
`gemini-2.5-pro`	Most capable Gemini
`gemini-2.5-flash`	Fast, cheap Gemini
`gemini-2.0-flash`	Older fast model, still available
`text-embedding-004`	Gemini embeddings

Override with GEMINI_MODELS=gemini-2.5-pro,gemini-2.5-flash.

Multiple instances

text

GEMINI_PROD_API_KEY=...
GEMINI_STAGING_API_KEY=...

Aurora exposes them as gemini-prod and gemini-staging.

Custom base URL

text

GEMINI_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai

This is the OpenAI-compatible endpoint. The native Vertex AI endpoint is not the same URL — use Vertex AI by setting GEMINI_BASE_URL to your regional Vertex endpoint and a service-account token as GEMINI_API_KEY.

Gotchas

Safety settings — Gemini has built-in safety filters. Aurora does not pass through custom safety_settings from the OpenAI format yet; rely on Aurora Guardrails for content policy.
Grounding / search — Google's grounding and tools.google_search parameters are not exposed through the OpenAI-compatible endpoint. Use provider passthrough at /p/gemini/* for native Gemini tools.
Image generation — Imagen is not available on the OpenAI-compatible endpoint. Use passthrough.
System prompt — Gemini treats the first system-style message as a system instruction. Aurora's guardrail transforms apply normally.
Prompt caching — Gemini's implicit cache is enabled on supported models; Aurora surfaces cache token counts in usage.

Sample request

text

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $AURORA_MASTER_KEY" \
  -d '{
    "model": "gemini/gemini-2.5-flash",
    "messages": [{"role": "user", "content": "Hello from Aurora"}]
  }'

Providers Overview
Passthrough API — for native Gemini features not in the OpenAI surface

Model

Use

gemini-2.5-pro

Most capable Gemini

gemini-2.5-flash

Fast, cheap Gemini

gemini-2.0-flash

Older fast model, still available

text-embedding-004

Gemini embeddings

Gotchas

Safety settings — Gemini has built-in safety filters. Aurora does not pass through custom safety_settings from the OpenAI format yet; rely on Aurora Guardrails for content policy.

Grounding / search — Google's grounding and tools.google_search parameters are not exposed through the OpenAI-compatible endpoint. Use provider passthrough at /p/gemini/* for native Gemini tools.

Image generation — Imagen is not available on the OpenAI-compatible endpoint. Use passthrough.

System prompt — Gemini treats the first system-style message as a system instruction. Aurora's guardrail transforms apply normally.

Prompt caching — Gemini's implicit cache is enabled on supported models; Aurora surfaces cache token counts in usage.

curl http://localhost:8080/v1/chat/completions \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $AURORA_MASTER_KEY" \ -d '{ "model": "gemini/gemini-2.5-flash", "messages": [{"role": "user", "content": "Hello from Aurora"}] }'

Gemini

Activate

Default models

Multiple instances

Custom base URL

Gotchas

Sample request

Related

Gemini

Activate

Default models

Multiple instances

Custom base URL

Gotchas

Sample request

Related