[ENHANCEMENT] Add Nebius AI as codebase indexing provider for cost-effective embeddings

[shariqriazz]

Problem (one or two sentences)

Users need more affordable embedding options for codebase indexing. Current providers cost
0.13
−
0.13−0.18 per million tokens, which can be expensive for large codebases.

Context (who is affected and when)

Affects all users who enable codebase indexing, especially those with large codebases or frequent re-indexing needs. Cost becomes significant during initial indexing and ongoing maintenance.

Desired behavior (conceptual, not technical)

Add Nebius AI as a provider option offering Qwen/Qwen3-Embedding-8B at $0.01 per million tokens - significantly cheaper than existing options while maintaining high quality (ranked #2 on MTEB Leaderboard, just below gemini-embedding-001).

Constraints / preferences (optional)

• Provider: Nebius AI
• API Endpoint: https://api.studio.nebius.com/v1/
• Model: Qwen/Qwen3-Embedding-8B
• Embedding dimensions: 4,096
• Rate limits: 600,000 TPM (tokens per minute), 10,000 RPM (requests per minute)
• Model specs: 7B parameters, 32K max tokens
• Pricing: 0.01 USD per 1M tokens (vs 0.13-0.18 USD per 1M for current providers)
• MTEB score: 70.58 (ranked Fix vscode compatibility issue #2, vs gemini-embedding-001 at 68.37)
• Should integrate similarly to existing providers (Gemini, OpenAI, Mistral, etc.)
• Maintain consistent user experience with current provider selection flow

Request checklist

• I've searched existing Issues and Discussions for duplicates
• This describes a specific problem with clear context and impact

Roo Code Task Links (optional)

N/A

Acceptance criteria (optional)

Given a user wants to configure codebase indexing
When they open the provider selection
Then Nebius AI should appear as an available option
And selecting it should allow configuration with API key
And indexing should work with Qwen/Qwen3-Embedding-8B model at endpoint https://api.studio.nebius.com/v1/
And embeddings should have 4,096 dimensions
And cost should be $0.01/1M tokens as advertised

Proposed approach (optional)

Follow the existing embedder pattern in src/services/code-index/embedders/ - similar to how Gemini, Mistral, and OpenAI are implemented. The provider should support the standard embedding interface with proper rate limiting and error handling. Configure with base URL https://api.studio.nebius.com/v1/ and model identifier Qwen/Qwen3-Embedding-8B.

Trade-offs / risks (optional)

New provider means additional maintenance overhead
Need to verify Nebius AI API stability and reliability
Should validate actual performance matches MTEB leaderboard claims
Confirm API compatibility with OpenAI-style endpoints

[roomote]

I'm exploring a potential fix for adding Nebius AI as an embedding provider and will open a PR to share progress.

[shariqriazz]

@roomote-agent please also consider the Rate limits: 600,000 TPM (tokens per minute), 10,000 RPM (requests per minute)

[roomote]

Thank you @shariqriazz for the clarification! I'll ensure the Nebius AI implementation includes proper rate limiting with 600,000 TPM and 10,000 RPM as specified. I'll incorporate this into the embedder implementation to prevent hitting these limits.

[roomote]

@shariqriazz I have implemented the rate limiting for Nebius AI as requested (600,000 TPM and 10,000 RPM).

The implementation includes:

• Sliding window rate limiting that tracks both token usage and request count
• Automatic window reset after 60 seconds
• Graceful handling with wait time calculation when limits are exceeded
• Comprehensive test coverage for rate limit scenarios

PR #8591 is ready for review: #8591

All CI checks have passed successfully. The rate limiting ensures the embedder respects both the token and request limits you specified. Feedback welcome!

[hannesrudolph]

For fringe options like this I advise simply using the OpenAI compatible option.