What Embedding Models Does AstrBot Support for Its Knowledge Base?

AstrBot supports OpenAI and Google Gemini embedding models out of the box, defaulting to text-embedding-3-small and gemini-embedding-exp-03-07 respectively.

AstrBot is an extensible AI chatbot framework that includes a knowledge base system for storing and retrieving vectorized text representations. Understanding what embedding models AstrBot supports is essential for configuring accurate semantic search and retrieval-augmented generation (RAG) capabilities. The framework provides built-in adapters for industry-leading embedding providers through a unified EmbeddingProvider interface.

Built-In Embedding Providers in AstrBot

AstrBot ships with two production-ready embedding providers that implement the abstract EmbeddingProvider class defined in astrbot/core/provider/provider.py. Both providers register themselves with the @register_provider_adapter decorator, making them available to the ProviderManager in astrbot/core/provider/manager.py.

OpenAI Embedding Provider

The OpenAI Embedding Provider connects to OpenAI or Azure OpenAI API endpoints to generate embeddings. According to the source code in astrbot/core/provider/sources/openai_embedding_source.py, this provider defaults to the text-embedding-3-small model when no specific model is configured.

Key configuration parameters:

• embedding_model: Model identifier (default: text-embedding-3-small, supports text-embedding-3-large or custom Azure deployments)
• embedding_api_key: Authentication key for the OpenAI API
• embedding_api_base: Optional endpoint URL for Azure OpenAI or proxy configurations
• embedding_dimensions: Vector dimensionality (default: 1024)

The provider calls client.embeddings.create() with the specified model and passes the dimensions parameter when get_dim() returns a non-None value.

Google Gemini Embedding Provider

The Google Gemini Embedding Provider interfaces with Google's Gemini API to generate vector representations. As implemented in astrbot/core/provider/sources/gemini_embedding_source.py, this provider defaults to the gemini-embedding-exp-03-07 model.

Key configuration parameters:

• embedding_model: Model identifier (default: gemini-embedding-exp-03-07)
• embedding_api_key: Google AI Studio API key
• embedding_dimensions: Vector dimensionality (default: 768)

This provider utilizes client.models.embed_content() with an EmbedContentConfig object that specifies the desired output dimensionality. The GeminiEmbeddingProvider.get_dim() method returns 768 by default, matching the Gemini embedding specification.

Configuring Embedding Models in AstrBot

Embedding configuration resides in the global configuration schema defined in astrbot/core/config/default.py (approximately lines 1610–1650). To activate an embedding provider for your knowledge base, specify the provider ID and model parameters in your configuration file.

knowledgebase:
  embedding_provider_id: openai_embedding
  embedding_provider_config:
    embedding_api_key: sk-your-openai-key
    embedding_model: text-embedding-3-large
    embedding_dimensions: 1536
    embedding_api_base: https://api.openai.com/v1

For Google Gemini:

knowledgebase:
  embedding_provider_id: gemini_embedding
  embedding_provider_config:
    embedding_api_key: your-gemini-api-key
    embedding_model: gemini-embedding-exp-03-07
    embedding_dimensions: 768

Using Embedding Providers Programmatically

Both providers implement the EmbeddingProvider abstract class, exposing get_embedding() for single texts and get_embeddings() for batches. The ProviderManager in astrbot/core/provider/manager.py handles instantiation and dependency injection.

from astrbot.core.provider.manager import ProviderManager

async def generate_embeddings():
    # Obtain the provider manager instance

    prov_mgr = ProviderManager.get_instance()
    
    # Retrieve the OpenAI embedding provider

    openai_ep = await prov_mgr.get_provider_by_id("openai_embedding")
    
    # Generate embedding for a single text

    text = "AstrBot is an extensible AI chatbot framework."
    vector = await openai_ep.get_embedding(text)
    print(f"Generated vector with dimension: {len(vector)}")
    
    # Batch processing

    texts = ["Hello world", "Knowledge base retrieval"]
    vectors = await openai_ep.get_embeddings(texts)
    return vectors

The KnowledgeBaseHelper class in astrbot/core/knowledge_base/kb_helper.py orchestrates these calls to index documents and perform similarity searches against the vector store.

Summary

• AstrBot supports OpenAI and Google Gemini embedding providers out of the box through the EmbeddingProvider interface.
• Default models are text-embedding-3-small (OpenAI, 1024 dimensions) and gemini-embedding-exp-03-07 (Gemini, 768 dimensions).
• Configuration occurs in astrbot/core/config/default.py with keys embedding_model, embedding_api_key, and embedding_api_base.
• Provider implementations reside in astrbot/core/provider/sources/openai_embedding_source.py and astrbot/core/provider/sources/gemini_embedding_source.py.
• Both providers expose identical async methods get_embedding() and get_embeddings() for single and batch processing.

Frequently Asked Questions

Can I use custom embedding models with AstrBot?

Yes, you can specify any model name supported by the underlying API in the embedding_model configuration field. For OpenAI, this includes text-embedding-3-large or custom Azure OpenAI deployment names. For Gemini, you can use newer experimental models as they become available in the Google AI Studio API.

What is the default embedding dimension for AstrBot providers?

The OpenAI provider defaults to 1024 dimensions when using text-embedding-3-small, while the Gemini provider defaults to 768 dimensions for gemini-embedding-exp-03-07. You can override these via the embedding_dimensions configuration key, though the underlying service must support the requested dimensionality.

How do I switch between OpenAI and Gemini embeddings?

Change the embedding_provider_id in your knowledge base configuration to either openai_embedding or gemini_embedding. After switching, you should rebuild your knowledge base vectors using the rebuild_vectors() method in KnowledgeBaseHelper to ensure all documents are embedded with the new model.

Where are the embedding provider implementations located?

The concrete implementations are located in astrbot/core/provider/sources/openai_embedding_source.py for OpenAI and astrbot/core/provider/sources/gemini_embedding_source.py for Gemini. The abstract interface they implement is defined in astrbot/core/provider/provider.py, and the registration logic resides in astrbot/core/provider/manager.py.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}