[BUG] MCP Server: cross_encoder/reranker not configurable, hardcoded to OpenAI — requires OPENAI_API_KEY even with non-OpenAI providers #1393

Bug Description

The MCP server does not pass a cross_encoder argument when constructing the Graphiti client. This causes Graphiti.__init__ to always fall back to OpenAIRerankerClient(), which requires OPENAI_API_KEY. This makes it impossible to run the MCP server without an OpenAI API key, even when using Anthropic/Gemini/Groq for both LLM and embeddings.

Steps to Reproduce

Configure the MCP server to use non-OpenAI providers for both LLM and embeddings:

# config.yaml
llm:
  provider: "anthropic"
  model: "claude-haiku-4-5-20251001"
  temperature: 0.0
  providers:
    anthropic:
      api_key: "${ANTHROPIC_API_KEY}"

embedder:
  provider: "gemini"
  model: "gemini-embedding-001"
  providers:
    gemini:
      api_key: "${GOOGLE_API_KEY}"

database:
  provider: "falkordb"

Start the MCP server. Both LLM and embedder clients are created successfully, but initialization crashes when Graphiti.__init__ tries to create the default OpenAIRerankerClient.

Expected Behavior

The MCP server should allow configuring the cross-encoder/reranker provider, or at minimum select a matching reranker based on the configured LLM provider.

Actual Behavior

The server crashes with:

openai.OpenAIError: The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable

Traceback:

File "/app/mcp/src/graphiti_mcp_server.py", line 225, in initialize
    self.client = Graphiti(
File "/app/mcp/.venv/lib/python3.11/site-packages/graphiti_core/graphiti.py", line 224, in __init__
    self.cross_encoder = OpenAIRerankerClient()
File "/app/mcp/.venv/lib/python3.11/site-packages/graphiti_core/cross_encoder/openai_reranker_client.py", line 55, in __init__
    self.client = AsyncOpenAI(api_key=config.api_key, base_url=config.base_url)
openai.OpenAIError: The api_key client option must be set either by passing api_key to the client or by setting the OPENAI_API_KEY environment variable

Environment

• Graphiti Version: graphiti-core 0.28.2, MCP server 1.0.2 (Docker image zepai/knowledge-graph-mcp:1.0.2-graphiti-0.28.2)
• Python Version: 3.11
• Operating System: Docker (Debian-based)
• Database Backend: FalkorDB
• LLM Provider & Model: Anthropic claude-haiku-4-5-20251001

Installation Method

• Docker (zepai/knowledge-graph-mcp:1.0.2-graphiti-0.28.2)

Root Cause

graphiti_mcp_server.py lines 225-231 and 233-239 construct Graphiti(...) passing llm_client and embedder but not cross_encoder:

self.client = Graphiti(
    graph_driver=falkor_driver,
    llm_client=llm_client,
    embedder=embedder_client,
    max_coroutines=self.semaphore_limit,
    # cross_encoder is missing here
)

Graphiti.__init__ (graphiti.py line 142) accepts an optional cross_encoder parameter, but when it's not provided, it defaults to OpenAIRerankerClient() (line 224), which unconditionally requires OPENAI_API_KEY.

Workaround

Monkey-patch Graphiti.__init__ to inject an alternative reranker before the server starts. For example, for Gemini:

import graphiti_core.graphiti as _graphiti

_orig_init = _graphiti.Graphiti.__init__

def _patched_init(self, *args, cross_encoder=None, **kwargs):
    if cross_encoder is None:
        from graphiti_core.cross_encoder.gemini_reranker_client import GeminiRerankerClient
        cross_encoder = GeminiRerankerClient()
    _orig_init(self, *args, cross_encoder=cross_encoder, **kwargs)

_graphiti.Graphiti.__init__ = _patched_init

Possible Solution

graphiti-core already ships GeminiRerankerClient and BGERerankerClient alongside OpenAIRerankerClient. The MCP server could:

1. Add a reranker (or cross_encoder) section to the config schema, mirroring the existing llm/embedder pattern.
2. Or, as a simpler default, infer the reranker from the configured LLM provider (e.g., Anthropic/Gemini LLM → GeminiRerankerClient, OpenAI → OpenAIRerankerClient).
3. Pass the resulting cross_encoder to the Graphiti() constructor.

Related

• [BUG] - OpenAIRerankerClient / cross encoder #834 — OpenAIRerankerClient ignoring model config (closed, narrower scope)
• Question: How to setup the graphiti to use my OpenRouter and Voyage? #517 — User confusion about configuring non-OpenAI providers in MCP Docker setup