Gemini Provider

Configure Google’s Gemini models with massive context windows in Avante.nvim

Google’s Gemini models offer industry-leading context windows (up to 1M+ tokens) and powerful multimodal capabilities. Avante.nvim provides full support for Gemini’s API.

Quick Start

1. Get your API key

   Get an API key from Google AI Studio.

2. Set environment variable

   Add to your shell configuration:

   # Scoped (recommended)
   export AVANTE_GEMINI_API_KEY=your-api-key
   
   # Or global
   export GEMINI_API_KEY=your-api-key
   
3. Configure provider

   {
     "yetone/avante.nvim",
     opts = {
       provider = "gemini",
     },
   }
   

Configuration

Basic Configuration

providers = {
  gemini = {
    endpoint = "https://generativelanguage.googleapis.com/v1beta/models",
    model = "gemini-2.0-flash",
    timeout = 30000,
    context_window = 1048576, -- 1M+ tokens
    extra_request_body = {
      generationConfig = {
        temperature = 0.75,
      },
    },
  },
}

Available Models

Gemini 2.0 Flash (Recommended)

providers = {
  gemini = {
    model = "gemini-2.0-flash",
    context_window = 1048576,
  },
}

Environment Variables

Variable	Scoped Version	Purpose
GEMINI_API_KEY	AVANTE_GEMINI_API_KEY	API authentication

API Endpoint Structure

Gemini uses a unique endpoint structure:

endpoint = "https://generativelanguage.googleapis.com/v1beta/models"

-- Full URL is constructed as:
-- {endpoint}/{model}:streamGenerateContent?alt=sse&key={api_key}

Generation Configuration

Gemini uses generationConfig for model parameters:

extra_request_body = {
  generationConfig = {
    temperature = 0.75,
    topP = 0.95,
    topK = 40,
    maxOutputTokens = 8192,
    candidateCount = 1,
  },
}

Parameters

Parameter	Type	Default	Description
temperature	number	0.75	Controls randomness (0.0-1.0)
topP	number	0.95	Nucleus sampling threshold
topK	number	40	Top-k sampling parameter
maxOutputTokens	number	8192	Maximum response length
candidateCount	number	1	Number of response candidates

Tool Calling

Gemini uses a different format for function declarations:

-- Avante automatically converts tools to Gemini format:
{
  tools = {
    {
      functionDeclarations = {
        {
          name = "tool_name",
          description = "Tool description",
          parameters = {
            type = "object",
            properties = { ... },
            required = { ... },
          },
        },
      },
    },
  },
}

Function Responses

-- Function results are formatted as:
{
  functionResponse = {
    name = "tool_name",
    response = {
      name = "tool_name",
      content = { result_data },
    },
  },
}

Safety Settings

Gemini includes safety filters. If your prompt is blocked:

-- Response includes:
{
  promptFeedback = {
    blockReason = "SAFETY", -- or "RECITATION"
  },
}

To adjust safety settings, add to your configuration:

extra_request_body = {
  safetySettings = {
    {
      category = "HARM_CATEGORY_HARASSMENT",
      threshold = "BLOCK_MEDIUM_AND_ABOVE",
    },
    -- Add other categories as needed
  },
}

Finish Reasons

Gemini responses include finish reasons:

Reason	Meaning
STOP	Natural completion
MAX_TOKENS	Reached token limit
SAFETY	Blocked by safety filters
RECITATION	Blocked due to recitation
TOOL_CODE	Tool use requested

Vertex AI

For Google Cloud Vertex AI, use the vertex provider:

providers = {
  vertex = {
    endpoint = "https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/publishers/google/models",
    model = "gemini-1.5-flash-002",
    timeout = 30000,
    context_window = 1048576,
    extra_request_body = {
      generationConfig = {
        temperature = 0.75,
      },
    },
  },
}

Authentication

Vertex AI uses Google Cloud authentication:

# Set up gcloud CLI authentication
gcloud auth application-default login

# Or use a service account
export GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json

Advanced Features

ReAct Prompting

Enable ReAct-style prompting for better tool use:

providers = {
  gemini = {
    use_ReAct_prompt = true,
  },
}

Stop Sequences

Custom stop sequences for ReAct mode:

extra_request_body = {
  generationConfig = {
    stopSequences = { "</tool_use>" },
  },
}

Multimodal Input

Gemini supports image inputs:

-- Images are automatically formatted as:
{
  inline_data = {
    mime_type = "image/png",
    data = "base64_encoded_data",
  },
}

Troubleshooting

API Key Not Found

Ensure your API key is set:

echo $GEMINI_API_KEY
# or
echo $AVANTE_GEMINI_API_KEY

Get a key from Google AI Studio.

Prompt Blocked by Safety Filters

If your prompt is blocked:

1. Review the blockReason in the error
2. Adjust your prompt to be less sensitive
3. Configure safety settings (use with caution)

Vertex AI Authentication Failed

For Vertex AI:

1. Ensure gcloud is authenticated: gcloud auth list
2. Check project ID in endpoint URL
3. Verify service account permissions

Rate Limit Errors

Gemini has generous quotas, but if you hit limits:

1. Check quota at Google Cloud Console
2. Request quota increase
3. Implement request throttling

Best Practices

Model Selection

• Gemini 2.0 Flash: Latest, best performance
• Gemini 1.5 Pro: Maximum capability
• Gemini 1.5 Flash: Fastest responses

Context Window

• 1M+ token context is unique to Gemini
• Great for large codebases
• Entire files can fit in context

Temperature Settings

• 0.0-0.3: Focused, consistent
• 0.4-0.7: Balanced (recommended)
• 0.8-1.0: Creative, varied

Safety

• Default filters are moderate
• Adjust only when necessary
• Review blocked content carefully

Example Configurations

Standard Setup

{
  provider = "gemini",
  providers = {
    gemini = {
      model = "gemini-2.0-flash",
      timeout = 30000,
      extra_request_body = {
        generationConfig = {
          temperature = 0.7,
          maxOutputTokens = 8192,
        },
      },
    },
  },
}

Related Resources

• Gemini API Documentation
• Google AI Studio
• Vertex AI Documentation
• Provider Overview