feat(mcp): add Prompts API and ragflow_list_datasets tool for intelligent retrieval routing

[Louisym]

Summary

This PR makes the RAGFlow MCP Server "white-box" to connected LLM clients by implementing
the MCP Prompts specification and adding a dataset discovery tool.

Before: clients saw a single opaque ragflow_retrieval tool with no guidance on which
knowledge bases exist, how to tune parameters, or how to recover from failures.

After: clients can request a live SOP prompt (ragflow_retrieval_skill) that injects:

• A real-time snapshot of all available knowledge bases (id, name, description, doc/chunk counts)
• Parameter tuning guide for RAGFlow's hybrid retrieval (Dense + BM25 + RRF)
• Routing decision rules (precise KB targeting vs. global fallback)
• Self-healing rules (empty results → lower threshold; 401 → check API key; 500 → retry once)

Changes

mcp/server/server.py

• New tool: ragflow_list_datasets — lets the LLM proactively refresh the KB list and
  get full metadata (name, document_count, chunk_count, embedding_model, language)
• New: @app.list_prompts() — exposes ragflow_retrieval_skill prompt with optional
  intent argument (precise / broad / auto)
• New: @app.get_prompt() — dynamically assembles a 4-module SOP prompt with a live
  KB snapshot on every call
• Refactor: list_datasets() — extended return fields and extracted shared helpers
  to eliminate code duplication; also populates the dataset metadata cache as a side effect

test/benchmark/mcp_skill_eval.py (new)

• A/B benchmark script comparing global search (baseline) vs. precise KB routing (with Skill)
• Integrates ragas (context_precision, context_recall) with OpenAI gpt-4o as judge LLM
• Reuses existing test/benchmark/ infrastructure (HttpClient, dataset helpers, metrics)

test/benchmark/wixqa_eval.py (new)

• WixQA-based benchmark using Wix/WixQA
  (200 expert-written + 200 simulated customer support Q&A pairs)
• Two-KB routing strategy: questions routed to the KB owning their source articles

Benchmark Results

Evaluated on 100 questions from Wix/WixQA
across 2 knowledge bases. Judged by gpt-4o via ragas.

Metric	Baseline (Both KBs)	With Routing	Delta
Context Precision	0.543	0.554	+2.0%
Context Recall	0.287	0.300	+4.5%
Avg Latency (ms)	403.1	366.2	-9.2%
Errors	0	0	—

Latency percentiles:

Stat	Baseline	Routing
p50	357.0 ms	338.4 ms
p90	516.4 ms	479.1 ms
p95	632.9 ms	531.2 ms

Routing consistently reduces noise (precision ↑), improves information coverage (recall ↑),
and cuts latency at every percentile by eliminating unnecessary KB searches.

Note: gains are intentionally conservative — the two WixQA KBs share the same domain
(Wix support articles), making routing harder than typical multi-domain deployments.
In heterogeneous KB setups (e.g., HR + Legal + Engineering), the precision delta is
expected to be significantly larger.

Test Plan

• Start MCP Server in self-host mode, connect via MCP Inspector
• Verify list_prompts returns ragflow_retrieval_skill with intent argument
• Verify get_prompt returns a prompt containing current KB list
• Verify intent=precise vs intent=broad produces different threshold guidance
• Verify ragflow_list_datasets tool returns full metadata JSON array
• Run benchmark: OPENAI_API_KEY=xxx uv run python test/benchmark/wixqa_eval.py --base-url http://127.0.0.1:9380 --api-key ragflow-xxx