feat(mcp): add configurable reranker/cross-encoder support

[sgaluza]

Summary

Add configurable reranker (cross-encoder) support to MCP server for 2-stage retrieval:

1. Embeddings find candidate passages (fast, approximate)
2. Reranker scores each passage against query (slower, but more precise)

Changes

• mcp_server/src/config/schema.py: Add RerankerConfig and RerankerProvidersConfig
• mcp_server/src/services/factories.py: Add RerankerFactory with OpenAI and Gemini support

Configuration

Environment Variables

# Enable OpenAI reranker (default)
RERANKER__PROVIDER=openai
RERANKER__MODEL=gpt-4.1-nano
RERANKER__PROVIDERS__OPENAI__API_KEY=sk-...

# Or use Gemini
RERANKER__PROVIDER=gemini
RERANKER__MODEL=gemini-2.5-flash
RERANKER__PROVIDERS__GEMINI__API_KEY=...

# Disable reranker
RERANKER__PROVIDER=none

YAML Config

reranker:
  provider: openai  # openai, gemini, none, disabled
  model: gpt-4.1-nano
  providers:
    openai:
      api_key: ${OPENAI_API_KEY}
    gemini:
      api_key: ${GOOGLE_API_KEY}

Graceful Degradation

If graphiti-core version doesn't include cross_encoder module, the factory will raise a clear error:

OpenAI Reranker not available in current graphiti-core version

Why gpt-4.1-nano as default?

Reranker performs simple True/False classification per passage. The task is lightweight, so:

• gpt-4.1-nano — faster, good for high-throughput
• gpt-5-nano — 2x cheaper input ($0.05 vs $0.10/1M), but has reasoning overhead

Users can choose based on their latency vs cost preferences.

---

Built OnSteroids