AI Assistant

Doccupine supports AI integration to enhance your documentation experience. You can use OpenAI, Anthropic, or Google Gemini to power AI features in your documentation site. The AI assistant uses your documentation content as context, allowing users to ask questions about your docs and receive accurate answers based on the documentation.

Opening the Assistant

Once a provider is configured, an assistant button appears in the header. You can open and close the assistant from anywhere on the site by pressing Command + I (Ctrl + I on Windows). Press the same shortcut again to dismiss it.

Setup

To enable AI features, create an .env file in the directory where your website is generated. By default, this is the nextjs-app/ directory.

Configuration

Create an .env file with the following configuration options:

# LLM Provider Configuration
# Choose your preferred LLM provider: openai, anthropic, or google
LLM_PROVIDER=openai

# API Keys (set the one matching your provider)
OPENAI_API_KEY=your_openai_api_key_here
ANTHROPIC_API_KEY=your_anthropic_api_key_here
GOOGLE_API_KEY=your_google_api_key_here

# Optional: Require a bearer token for direct /api/rag requests on a public site
# RAG_API_KEY=generate-a-long-random-secret

# Optional: Override default chat model (see your provider's docs for available models)
# LLM_CHAT_MODEL=your-model-id

# Optional: Override default embedding model (see your provider's docs for available models)
# Note: Anthropic doesn't provide embeddings, will fallback to OpenAI
# LLM_EMBEDDING_MODEL=your-embedding-model-id

# Optional: Set temperature (0-1, default: 0)
# LLM_TEMPERATURE=0

# Optional: Embedding dimensions for the prebuilt search index (default: 512)
# Doc vectors are truncated to this many dimensions and stored as int8 to keep
# the search index small. Lower = smaller index, slightly lower recall.
# Rebuild after changing this.
# LLM_EMBEDDING_DIMS=512

Provider Selection

Set LLM_PROVIDER to one of the following values:

• openai - Use OpenAI's models
• anthropic - Use Anthropic's models
• google - Use Google's models

API Keys

You need to set the API key that matches your chosen provider:

• For OpenAI: Set OPENAI_API_KEY
• For Anthropic: Set ANTHROPIC_API_KEY
• For Google: Set GOOGLE_API_KEY

Protecting model spend

Public documentation intentionally leaves the browser assistant available by default. For a public deployment where only server-to-server clients should call the paid RAG endpoint, set RAG_API_KEY and require Authorization: Bearer <key> on /api/rag.

Using Anthropic with OpenAI

If you want to use Anthropic as your LLM provider, you must also have an OpenAI API key set. Here's why:

The Situation

Anthropic (Claude) does not provide an embeddings API. They only offer chat/completion models, not text embeddings.

Your RAG (Retrieval-Augmented Generation) system has two components:

• Chat/Completion - Generates answers, works with Anthropic.
• Embeddings - Creates vector representations of text for search, Anthropic doesn't provide this.

When using Anthropic as your LLM_PROVIDER, Doccupine will use Anthropic for chat/completion tasks, but will automatically fallback to OpenAI for embeddings. This means you need both API keys configured:

LLM_PROVIDER=anthropic
ANTHROPIC_API_KEY=your_anthropic_api_key_here
OPENAI_API_KEY=your_openai_api_key_here

This hybrid approach allows you to leverage Anthropic's powerful chat models while still having access to embeddings functionality through OpenAI.

Default models

Optional Settings

Chat Model

Override the default chat model by uncommenting and setting LLM_CHAT_MODEL. You can use any available model from your chosen provider. For a complete list of available models, refer to the official documentation:

• OpenAI Models
• Anthropic Models
• Google Gemini Models

Embedding Model

Override the default embedding model by uncommenting and setting LLM_EMBEDDING_MODEL. For a complete list of available embedding models, refer to the official documentation:

• OpenAI Embeddings
• Google Gemini Embeddings
• Anthropic: Anthropic doesn't provide embeddings. If you use Anthropic as your provider, Doccupine will fallback to OpenAI for embeddings.

Temperature

Control the randomness of AI responses by setting LLM_TEMPERATURE to a value between 0 and 1:

• 0 - More deterministic and focused responses (default)
• 1 - More creative and varied responses

Embedding Dimensions

Control the size of the prebuilt search index by setting LLM_EMBEDDING_DIMS (default: 512). Doc vectors are truncated to this many dimensions and stored as int8, keeping services/mcp/docs-index.json small so large doc sets don't stall the AI chat on serverless cold starts.

• Lower values produce a smaller index with slightly lower search recall.
• Values greater than or equal to the model's native dimension keep full precision.

This value must match between build time and runtime. A mismatch forces Doccupine to re-embed your docs at runtime, so rebuild or redeploy your site after changing it.