feat(mcp): use a provider-matched cross-encoder to avoid a hard OpenAI dependency

[abouchard11]

Problem

The MCP server builds an LLM client and an embedder client from config, but never passes a cross_encoder to Graphiti(). When Graphiti() receives no cross-encoder it hard-defaults to OpenAIRerankerClient(), which requires OPENAI_API_KEY.

The result: a deployment configured entirely for a non-OpenAI provider (e.g. Gemini LLM + Gemini embedder) still fails to start without an OpenAI key, purely because of the reranker default.

Fix

Add a CrossEncoderFactory that mirrors the existing LLMClientFactory / EmbedderFactory:

• for the gemini provider it returns a GeminiRerankerClient (already shipped in graphiti-core) built from the configured Gemini API key;
• every other provider returns None, which preserves the current default (graphiti-core applies OpenAIRerankerClient()).

The result is wired into both the FalkorDB and Neo4j Graphiti() constructors.

Compatibility

No behavior change for existing OpenAI users: unmatched providers return None, so graphiti-core keeps applying its OpenAIRerankerClient() default exactly as before. An all-Gemini deployment can now start without an OPENAI_API_KEY.

Verification

• Both changed files byte-compile.
• Imports resolve in the providers extra (HAS_GEMINI_RERANKER = True).
• gemini → constructs a GeminiRerankerClient; anthropic/other → returns None (default preserved).

Scope

Two files, additive:

• mcp_server/src/services/factories.py — new CrossEncoderFactory + a guarded GeminiRerankerClient import.
• mcp_server/src/graphiti_mcp_server.py — build the cross-encoder via the factory and pass it into both Graphiti() constructors.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[zep-cla-assistant]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[abouchard11]

I have read the CLA Document and I hereby sign the CLA behalf on myself, e-mail: alex@midnightdev.dev

[abouchard11]

I have read the CLA Document and I hereby sign the CLA

[Naseem77]

@abouchard11 Tested the factory: Gemini gets a proper reranker, but Anthropic/other providers return None and still fall back to the OpenAI reranker, so the hard OPENAI_API_KEY requirement remains for them. #1637 covers all providers with a local BGE fallback and tests, probably best to consolidate there.

[abouchard11]

Consolidating into #1637 as @Naseem77 suggested — it's a superset of this change (provider inference including the embedder, local BGE fallback, tests) and covers the Anthropic case flagged above. I've tested #1637 against my production Neo4j + all-Gemini stack and left a review with findings there. Closing.