Supported Embedding Providers in code-review-graph: Configuration Guide

The code-review-graph repository supports four pluggable embedding providers—local (sentence-transformers), Google Gemini, MiniMax, and OpenAI-compatible endpoints—each activated via environment variables and the get_provider() factory in code_review_graph/embeddings.py.

code-review-graph implements a flexible embedding layer to power semantic code search across your repositories. The system abstracts vector generation through a provider pattern defined in code_review_graph/embeddings.py, allowing you to switch between offline local models and cloud-based APIs without changing core logic. Understanding these supported embedding providers and their configuration options is essential for optimizing performance and maintaining data privacy.

Overview of Embedding Providers

The embedding subsystem centers on the get_provider() factory function (lines 74-86 in code_review_graph/embeddings.py), which validates names against the _VALID_PROVIDERS set and returns concrete implementations:

• LocalEmbeddingProvider: Offline inference using sentence-transformers with no API keys required.
• GoogleEmbeddingProvider: Cloud-based Gemini embeddings via the google-generativeai client.
• MiniMaxEmbeddingProvider: MiniMax's proprietary embo-01 1536-dimensional embeddings.
• **OpenAIEmbeddingProvider: Any OpenAI-compatible /v1/embeddings` endpoint, including Azure OpenAI and self-hosted gateways.

Configuring the Local Provider

The LocalEmbeddingProvider enables fully offline embedding generation, making it ideal for air-gapped environments. It lazily imports sentence_transformers and defaults to the all-MiniLM-L6-v2 model unless overridden via CRG_EMBEDDING_MODEL.

export CRG_EMBEDDING_MODEL=all-MiniLM-L6-v2
python -m code_review_graph.main --provider local

If the sentence_transformers package is missing, the provider is unavailable and the system restricts you to cloud options only. No API keys or network access are required for this provider.

Configuring the Google Gemini Provider

The GoogleEmbeddingProvider connects to Google's Gemini embedding API and requires the google-generativeai Python package. The constructor validates dependencies and raises explicit errors if the package is absent.

Set your configuration via environment variables:

export GOOGLE_API_KEY=YOUR_GEMINI_KEY
export CRG_EMBEDDING_MODEL=gemini-embedding-001  # optional override

python -m code_review_graph.main --provider google

This provider is categorized as a cloud provider in the CLOUD_PROVIDERS set (lines 28-30).

Configuring the MiniMax Provider

The MiniMaxEmbeddingProvider interfaces with MiniMax's embo-01 embedding service, producing 1536-dimensional vectors suitable for semantic similarity tasks.

Required configuration:

export MINIMAX_API_KEY=YOUR_MINIMAX_KEY
python -m code_review_graph.main --provider minimax

The API key is stored on the provider instance and used for every HTTP request to the MiniMax API.

Configuring OpenAI-Compatible Providers

The OpenAIEmbeddingProvider supports any endpoint implementing the OpenAI /v1/embeddings specification, including official OpenAI, Azure OpenAI Service, and custom reverse proxies.

Required environment variables (validated in lines 22-27):

• CRG_OPENAI_API_KEY: Authentication token
• CRG_OPENAI_BASE_URL: Endpoint URL (e.g., https://api.openai.com/v1)
• CRG_OPENAI_MODEL: Model identifier (e.g., text-embedding-3-small)

Optional overrides:

• CRG_OPENAI_DIMENSION: Vector dimensionality
• CRG_OPENAI_BATCH_SIZE: Request batching limits

export CRG_OPENAI_API_KEY=sk-...
export CRG_OPENAI_BASE_URL=https://api.openai.com/v1
export CRG_OPENAI_MODEL=text-embedding-3-small
export CRG_OPENAI_DIMENSION=1536
python -m code_review_graph.main --provider openai

The constructor raises ValueError listing any missing required variables.

Cloud Provider Safety Controls

All cloud providers (google, minimax, openai) emit a one-time warning to stderr before making outbound requests, as defined by the CLOUD_PROVIDERS constant. To acknowledge data transmission to third-party APIs and suppress this warning, set:

export CRG_ACCEPT_CLOUD_EMBEDDINGS=1

This safety mechanism is implemented in lines 45-52 of code_review_graph/embeddings.py.

Summary

• Four providers available: Local (sentence-transformers), Google (Gemini), MiniMax, and OpenAI-compatible, each with distinct latency and privacy characteristics.
• Configuration method: Environment variables for API keys and model selection, parsed by get_provider() in code_review_graph/embeddings.py.
• Offline capability: The local provider requires no API keys but depends on the sentence_transformers package.
• Safety features: Cloud providers warn before API calls unless CRG_ACCEPT_CLOUD_EMBEDDINGS=1 is exported.

Frequently Asked Questions

Which embedding provider works without an internet connection?

The local provider using LocalEmbeddingProvider operates entirely offline via sentence-transformers. You only need to install the Python package and optionally set CRG_EMBEDDING_MODEL to specify a model like all-MiniLM-L6-v2. No API keys or network connectivity are required.

How do I migrate from OpenAI to Google Gemini embeddings?

Change your environment configuration by unsetting CRG_OPENAI_* variables and exporting GOOGLE_API_KEY with your Gemini API key. Then specify --provider google when running code_review_graph.main. The get_provider() factory automatically instantiates GoogleEmbeddingProvider based on the provider string passed to it.

What triggers the cloud embedding warning in code-review-graph?

Any provider contained in the CLOUD_PROVIDERS set (google, minimax, openai) triggers a stderr warning before the first API request to alert you about data leaving your local network. Set CRG_ACCEPT_CLOUD_EMBEDDINGS=1 to acknowledge this transmission and suppress the warning message.

Can I use Azure OpenAI Service instead of the standard OpenAI API?

Yes. The OpenAI-compatible provider supports any /v1/embeddings endpoint. Configure CRG_OPENAI_BASE_URL to your Azure OpenAI endpoint (e.g., https://YOUR_RESOURCE.openai.azure.com/openai/deployments/YOUR_DEPLOYMENT/) and provide the Azure API key in CRG_OPENAI_API_KEY. The system treats this as a standard OpenAI-compatible endpoint.

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