How to Configure MiniMax Embeddings for code-review-graph Semantic Search

To configure MiniMax embeddings for semantic search in code-review-graph, set the MINIMAX_API_KEY environment variable, optionally export CRG_ACCEPT_CLOUD_EMBEDDINGS=1 to acknowledge cloud data processing, and specify --embedding-provider minimax via CLI or provider="minimax" in Python.

The code-review-graph library supports multiple vector embedding backends for semantic code search, including the cloud-based MiniMax provider (embo-01). Configuring MiniMax embeddings enables high-quality 1536-dimensional vector representations of your codebase stored in a local SQLite database. This guide walks through the exact environment variables, CLI flags, and Python API calls required to enable MiniMax as implemented in code_review_graph/embeddings.py.

Prerequisites: API Key and Cloud Egress Acknowledgment

Before initializing the MiniMax provider, you must satisfy two security requirements hardcoded in the source.

Set the MiniMax API Key

The MiniMaxEmbeddingProvider class reads your API key from the MINIMAX_API_KEY environment variable. Without this variable, the provider cannot authenticate against the MiniMax endpoint at https://api.minimax.io/v1/embeddings.

export MINIMAX_API_KEY="your-minimax-api-key-here"

Accept Cloud Data Processing

Because MiniMax is a third-party cloud service, the library implements a safety check in the _warn_cloud_egress function within code_review_graph/embeddings.py. Unless you set CRG_ACCEPT_CLOUD_EMBEDDINGS=1, the library prints a one-time warning about sending code snippets to external APIs.

export CRG_ACCEPT_CLOUD_EMBEDDINGS=1

Configure MiniMax via Command Line Interface

The CLI entry point in code_review_graph/cli.py accepts an --embedding-provider argument that routes to the get_provider factory function. To use MiniMax for indexing your repository:

code-review-graph update \
    --embedding-provider minimax \
    --embedding-model embo-01 \
    --repo /path/to/your/repo

This command initializes a MiniMaxEmbeddingProvider instance with the embo-01 model, batches your source files (default 100 texts per batch), and stores the resulting 1536-dimensional vectors in the SQLite database.

Configure MiniMax Programmatically

For custom pipelines, instantiate EmbeddingStore with provider="minimax" or use the lower-level get_provider factory directly.

Using EmbeddingStore for Search

The EmbeddingStore class manages the SQLite persistence layer and vector search logic. When you specify provider="minimax", it internally calls get_provider("minimax") to construct the provider.

from code_review_graph.embeddings import EmbeddingStore

# Initialize with MiniMax provider

store = EmbeddingStore(
    db_path="mygraph.db",
    provider="minimax",
    model="embo-01"
)

# Perform semantic search

results = store.search("how to build a spring service", limit=10)
for qualified_name, score in results:
    print(f"{score:.3f}  {qualified_name}")

Direct Provider Access

If you need embeddings without SQLite storage, import get_provider to obtain a MiniMaxEmbeddingProvider instance:

from code_review_graph.embeddings import get_provider

provider = get_provider(provider="minimax")

# Batch embed documents (automatic batching at 100 items)

vectors = provider.embed(["def foo(): pass", "class Bar:"])

# Embed a single query

query_vec = provider.embed_query("function that does nothing")

MiniMax Implementation Details

Understanding the internal architecture helps debug configuration issues.

• Model Specification: The provider uses the embo-01 model (_MODEL constant) producing 1536-dimensional vectors (_DIMENSION constant) as defined in code_review_graph/embeddings.py.
• API Endpoint: Requests route to https://api.minimax.io/v1/embeddings (_ENDPOINT constant) via the _call_api method, which includes retry logic for resilience.
• Task Types: The provider distinguishes between indexing (task="db") and querying (task="query") when calling the MiniMax API, automatically set by embed() versus embed_query() methods.
• Integration Points: During indexing, embed_nodes calls self.provider.embed(texts) to populate the SQLite embeddings table. At query time, search uses self.provider.embed_query(query) to vectorize the search phrase before computing cosine similarity.

Summary

• Export MINIMAX_API_KEY to authenticate with the MiniMax cloud API endpoint.
• Set CRG_ACCEPT_CLOUD_EMBEDDINGS=1 to suppress the cloud egress warning implemented in _warn_cloud_egress.
• Use --embedding-provider minimax with the CLI or provider="minimax" with the EmbeddingStore constructor.
• The provider generates 1536-dimensional vectors using the embo-01 model with automatic batching of 100 texts per request handled by _call_api.
• Core implementation resides in code_review_graph/embeddings.py with CLI argument parsing in code_review_graph/cli.py.

Frequently Asked Questions

What embedding model does the MiniMax provider use?

The MiniMaxEmbeddingProvider hardcodes the embo-01 model, which produces 1536-dimensional vectors suitable for high-accuracy semantic code search. This is defined as the _MODEL constant in code_review_graph/embeddings.py.

Can I use MiniMax embeddings without storing them in SQLite?

Yes. Instead of using EmbeddingStore, call get_provider("minimax") to obtain a raw provider instance. You can then call provider.embed() for documents or provider.embed_query() for search queries without database persistence.

Why do I need to set CRG_ACCEPT_CLOUD_EMBEDDINGS?

This environment variable acknowledges that your code snippets will be sent to MiniMax's cloud API (https://api.minimax.io). The _warn_cloud_egress function in code_review_graph/embeddings.py enforces this opt-in to prevent accidental data leakage to third-party services.

Is there a limit to how many files I can embed at once?

The MiniMaxEmbeddingProvider automatically handles batching with a default size of 100 texts per API call. You can embed thousands of files; the provider will chunk them appropriately in the _call_api method to respect rate limits and payload size constraints.

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