fix: implement model-specific rate limiting for gemini-embedding-001

[roomote]

Fixes #5713

Problem

The gemini-embedding-001 model was hitting quota limits (HTTP 429) during indexing with "Batch Embed Content API requests" per-minute limit exceeded, while text-embedding-004 worked fine with the same API key.

Root Cause

Both Gemini models were using the same batching configuration:

• Same batch token limit (100,000 tokens)
• Same retry delays (500ms initial)
• No model-specific rate limiting

However, gemini-embedding-001 has stricter per-minute batch embedding limits compared to text-embedding-004.

Solution

Implemented model-specific rate limiting for gemini-embedding-001:

Changes Made

1. Added Gemini-specific constants:

   • GEMINI_EMBEDDING_001_MAX_BATCH_TOKENS = 20,000 (reduced from 100,000)
   • GEMINI_EMBEDDING_001_RETRY_DELAY_MS = 2,000 (increased from 500ms)
   • GEMINI_EMBEDDING_001_MAX_BATCH_SIZE = 10 (new batch size limit)

2. Enhanced OpenAICompatibleEmbedder:

   • Added configurable batch token limits
   • Added configurable retry delays
   • Added configurable batch size limits
   • Added inter-batch delays for models with stricter rate limits

3. Updated GeminiEmbedder:

   • Detects gemini-embedding-001 model and applies specific configuration
   • Uses default configuration for other models like text-embedding-004

Benefits

• Prevents quota exhaustion: Smaller batches and longer delays respect API limits
• Maintains performance: Only applies restrictions to gemini-embedding-001
• Backward compatible: No changes to existing text-embedding-004 behavior
• Configurable: Easy to adjust limits if Google changes quotas

Testing

• ✅ All existing tests pass
• ✅ Updated test expectations for new constructor signature
• ✅ Type checking passes
• ✅ Linting passes

This should resolve the indexing failures for users with gemini-embedding-001 while maintaining optimal performance for text-embedding-004.

---

Important

Implements model-specific rate limiting for gemini-embedding-001 in GeminiEmbedder and OpenAICompatibleEmbedder with new constants and tests.

• Behavior:
  • Implements model-specific rate limiting for gemini-embedding-001 in GeminiEmbedder and OpenAICompatibleEmbedder.
  • Adds inter-batch delays for models with stricter rate limits in OpenAICompatibleEmbedder.
• Constants:
  • Adds GEMINI_EMBEDDING_001_MAX_BATCH_TOKENS, GEMINI_EMBEDDING_001_RETRY_DELAY_MS, and GEMINI_EMBEDDING_001_MAX_BATCH_SIZE to constants/index.ts.
• Testing:
  • Updates gemini.spec.ts to test new constructor parameters and behavior for GeminiEmbedder.

^{This description was created by for 5f27712. You can customize this summary. It will automatically update as commits are pushed.}

[SatoshiReport]

Code Review Summary

I've reviewed the implementation for issue #5714 and the changes look excellent! Here's my analysis:

✅ Strengths

1. Targeted Solution: The fix specifically addresses gemini-embedding-001 rate limits without affecting text-embedding-004 performance
2. Clean Architecture: Model-specific configuration is handled elegantly through constructor parameters
3. Proper Constants: All magic numbers are extracted to named constants with clear documentation
4. Backward Compatibility: Existing behavior for text-embedding-004 remains unchanged
5. Test Coverage: Tests are updated to cover both model configurations
6. Inter-batch Delays: Smart addition of delays between batches for rate-limited models

🎯 Key Implementation Details

• Batch Size Reduction: 20K tokens vs 100K (5x smaller batches)
• Retry Delay Increase: 2000ms vs 500ms (4x longer delays)
• Batch Count Limit: Max 10 items per batch for gemini-embedding-001
• Smart Detection: Automatic model detection and configuration application

📋 Code Quality

• ✅ Follows fail-fast principles
• ✅ Clear, unambiguous naming conventions
• ✅ No magic numbers or hardcoded values
• ✅ Proper error handling maintained
• ✅ Test coverage updated appropriately

🚀 Ready for Merge

This implementation should resolve the quota exhaustion issues for gemini-embedding-001 users while maintaining optimal performance for other models. The solution is well-architected, properly tested, and follows the project's coding standards.

Recommendation: ✅ Approve and merge

[SatoshiReport]

Code Review: ✅ APPROVED

Excellent implementation of model-specific rate limiting for gemini-embedding-001!

🎯 Strengths:

• Targeted solution: Only affects the problematic model
• Clean architecture: Configurable parameters through constructor
• Proper constants: All magic numbers extracted with clear names
• Backward compatible: No breaking changes to existing functionality
• Test coverage: Updated tests for new behavior

🚀 Key improvements:

• Batch size: 20K tokens (vs 100K) - 5x reduction
• Retry delays: 2000ms (vs 500ms) - 4x increase
• Batch limits: Max 10 items per batch
• Inter-batch delays: Smart rate limiting for gemini-embedding-001

📋 Code Quality Assessment:

• ✅ Follows fail-fast principles with clear error handling
• ✅ Uses explicit, unambiguous naming conventions
• ✅ Implements targeted solution without scope creep
• ✅ Maintains clean architecture with single responsibility
• ✅ Has proper test coverage

This should resolve the quota exhaustion issues (HTTP 429 errors) while maintaining optimal performance for text-embedding-004.

Ready to merge! 🚀