feat: add OpenAI embedding backend for vector indexing

[rorybyrne]

Summary

Add an OpenAI embedding backend as an alternative to the local sentence-transformers model. This enables much faster bulk indexing at minimal cost.

Motivation

The current local embedding (MiniLM-L6-v2) takes ~30 seconds per batch of 64 records on a small Fly.io instance. For 250k records, this translates to ~75 hours of indexing time.

OpenAI's embedding API can process the same dataset in minutes for under $1:

• 250k records × ~150 tokens = ~37M tokens
• text-embedding-3-small: $0.02/1M tokens = ~$0.75 total
• text-embedding-3-large: $0.13/1M tokens = ~$5 total

Proposed Implementation

1. New backend class: OpenAIStorageBackend in osa/infrastructure/index/openai/

2. Config:

   class OpenAIBackendConfig(BackendConfig):
       api_key: str  # or use OPENAI_API_KEY env var
       model: str = "text-embedding-3-small"
       batch_size: int = 2048  # OpenAI supports up to 2048 inputs per request
       dimensions: int | None = None  # Optional dimensionality reduction

3. Backend implementation:

   • Use openai Python SDK (async client)
   • Batch requests (up to 2048 embeddings per API call)
   • Store in ChromaDB (same as current vector backend)
   • Handle rate limits with exponential backoff

4. Config selection: Allow choosing backend type in index config:

   indexes:
     vector:
       type: openai  # or "local" for sentence-transformers
       model: text-embedding-3-small

Alternatives Considered

• Voyage AI: Similar pricing, good quality, but OpenAI is more widely used
• Cohere: Slightly more expensive ($0.10/1M tokens)
• Larger Fly instance: More expensive than API costs for bulk indexing
• GPU instance: Overkill for this use case

Tasks

• Add openai to dependencies
• Create OpenAIBackendConfig
• Implement OpenAIStorageBackend with batching and rate limit handling
• Update config to support backend type selection
• Add integration test with mocked OpenAI responses
• Update documentation