Spaces:

---

Ekimetrics

/

ClimateQA-MCP

like 0

Running

Summary

• Adds mcp_server/logging.py, a self-contained module for persisting MCP tool calls to a private Hugging Face dataset (mirrors the Gradio chat logging pattern, but scoped to the MCP server).
• Validates the dataset repo at startup via async httpx; if env vars are missing, disabled, or the repo is unreachable, logging is turned off for the process with a single log line (no repeated failures on every tool call).
• Wires retrieve_data_mcp and retrieve_graphs to log input, structured output, and FastMCP session metadata (session_id, request_id, client_id, client_info).
• Documents new env vars in the README.

Log format

Each tool call is uploaded as mcp/{tool}_{timestamp}.json:

{
  "tool": "retrieve_data_mcp",
  "input": { "query": "..." },
  "output": { "documents": [...], "images": [...] },
  "session_id": "...",
  "request_id": "...",
  "client_id": "...",
  "client_info": { "name": "...", "version": "..." },
  "time": "...",
  "timestamp": "..."
}

The envelope is tool-agnostic — future tools (e.g. TTD) only need to call log_tool_call(...).

Configuration (Space secrets)
Variable Purpose
HF_LOGS_MCP_TOKEN
HF write token (fine-grained token scoped to the log dataset recommended)
HF_DATASET_REPO_MCP
e.g. Ekimetrics/climateqa-mcp-logs
MCP_LOG_TO_HF
false to disable (default: true)
The MCP Space can remain public; the token lives in Space secrets and the dataset stays private.

Design notes
Startup short-circuit: init_logging() runs in the lifespan before serving requests; _ready gates all uploads.
Non-blocking: uploads run in a background thread (asyncio.to_thread); tool responses are not awaited on HF I/O.
Decoupled from climateqa: no dependency on the sync Gradio logging path in climateqa/logging.py.