leeroopedia-mcp

Leeroopedia MCP

Leeroopedia is an ML/AI knowledge wiki. This skill teaches you when and how to call the Leeroopedia MCP tools so answers are grounded in documented best practices (not guesswork).

Default behavior

Use Leeroopedia MCP whenever the user asks anything that depends on ML/AI framework specifics or best practices (fine-tuning, post-training, inference serving, CUDA/Triton kernels, distributed training, RAG/agents, evaluation, config formats, API contracts, performance tuning).

If the question is purely general software engineering (no ML/AI-specific uncertainty), you may answer without tools.

Grounding policy

1. Prefer tool-grounded answers over memory. If you are not 100% sure about an ML/AI detail, call a tool.
2. Cite sources. When tools return [PageID] citations, preserve them in your final answer.
3. Expand citations when needed. If the user asks “show me the source”, “give the full reference”, or you need precise details, call get_page on the cited [PageID].
4. Use parallel searches. For ambiguous topics, call search_knowledge multiple times with different angles (faster and higher recall than a single broad query).

Tool selection guide

1) search_knowledge(query, context?)

Use when: you need documented facts: framework behavior, APIs, configs, conventions, design patterns, tradeoffs.
How to query well:

• Include framework + component + intent + constraints (model size, GPUs, latency/throughput target, memory limit, etc.)
• Ask narrow questions; do multiple calls instead of one giant query.
• Add context if you’re implementing a specific system.

Good query patterns

• “How does X work in Y (version)? What are the gotchas?”
• “What config fields are required for Z?”
• “Best practices for A given constraints B”

2) build_plan(goal, constraints?)

Use when: the user wants an end-to-end or multi-step implementation plan (pipelines, training runs, deployments, evaluations).
Output expectation: overview, key specs, numbered steps, validation criteria.