feat: Add output_dimension support for Voyage AI embeddings

[lagiosv]

Summary

This PR adds support for custom embedding dimensions when using Voyage AI's embedding models through the OpenAI-compatible API.

Problem

LightRAG's OpenAI embedding binding passes the dimensions parameter directly to the API. However, Voyage AI's OpenAI-compatible endpoint requires the dimension parameter to be passed as output_dimension via the extra_body parameter.

Without this fix:

• embedding_dim is configured (e.g., 2048)
• Voyage AI ignores the dimensions parameter
• Voyage AI returns 1024-dimensional vectors (default)
• Dimension mismatch errors occur with vector stores

Related issues: #2233, #2230, #2119

Solution

Detect Voyage AI by checking if base_url contains "voyageai.com" and route the dimension parameter appropriately:

• For Voyage AI: pass output_dimension via extra_body parameter
• For OpenAI and others: pass dimensions directly (existing behavior)

Changes

lightrag/llm/openai.py

# Add dimensions parameter only if embedding_dim is provided
if embedding_dim is not None:
    # Voyage AI requires 'output_dimension' via extra_body, not 'dimensions'
    is_voyage_ai = base_url and "voyageai.com" in base_url
    if is_voyage_ai:
        api_params["extra_body"] = {"output_dimension": embedding_dim}
    else:
        api_params["dimensions"] = embedding_dim

Testing

Tested with:

• Voyage AI voyage-3-large model
• embedding_dim=2048
• Qdrant vector store

Verified embeddings are correctly 2048-dimensional.

Backwards Compatibility

• No breaking changes
• Existing OpenAI configurations work unchanged
• Only affects Voyage AI when embedding_dim is explicitly set