RAG MCP LSP Ops

Overview

This skill defines the canonical operational workflow for RAG + MCP + LSP in this repository. Use it when the assistant needs reliable code-context retrieval, semantic navigation readiness, or full RAG rebuild. For contract governance, runbook observability and deep proposal generation, combine with:

• audit-contracts-v3-ops
• audit-runbook-observability
• audit-proposal-deep-triage

When specifically working with language-server features (definitions, completions, diagnostics), you can also pair this with the dedicated lsp-ops skill for finer semantic guidance.

When To Use

• User asks to rebuild/reindex RAG.
• MCP tools are unavailable or inconsistent.
• LSP tools need validation before deeper coding work.
• Assistant must prepare runtime (pm2, server, mcp) before coding tasks that depend on tool context.

Canonical Workflow

1. Preflight runtime.

• Check Ollama local: curl http://host.docker.internal:11434/api/version
• Ensure PM2 runtime: npm run daemon:start
• Wait endpoints: GET /health and GET /api/mcp on http://localhost:3008

2. Diagnose MCP.

• Run npm run mcp:diagnose
• Confirm required tools (rag_*, ollama_*, lsp_*) are listed.

3. Rebuild RAG from zero.

• Run npm run rag:rebuild:zero
• If needed, force profile full manually: npm run rag:index -- --profile full
• For scope control, use the same flags across rebuild/index/watch:
  • --docs-mode include|exclude|only
  • --include-glob / --exclude-glob (repeatable)
  • --max-file-bytes
• Recommended phased indexing:
  • Code/config first (no markdown): npm run rag:index -- --profile full --docs-mode exclude
  • Docs later: npm run rag:index -- --profile full --docs-mode only

4. Validate usable retrieval.

• Execute one rag_search smoke call over MCP (tools/call) and confirm non-error response.

5. Validate index freshness policy.

• Prefer index freshness within 24h for high-signal triage.
• If index_freshness_ms is stale, mark reduced confidence and recommend refresh/rebuild.

6. Continue coding flow.

• For normal coding queries use profile=core.
• Reserve profile=full for rebuild/reindex and wide investigation.

Tooling Policy

• Prefer MCP tools before broad local text search when task is semantic code understanding:
  • rag_search first for recall.
  • Then lsp_definition, lsp_references, lsp_hover, lsp_diagnostics.
• Daily coding order:
  • Symbol/API-level navigation: LSP (lsp_*) first.
  • Broad cross-file context: rag_search.
  • If embeddings are degraded: keep working in lexical mode temporarily, with diagnostics enabled.
• If MCP is down, recover via canonical preflight, then retry MCP tools.

Failure Handling

• mcp:diagnose failing:
  • Recheck PM2 status (npm run daemon:status)
  • Re-run npm run daemon:restart
  • Re-run npm run mcp:diagnose
• RAG health failing:
  • Confirm OLLAMA_LOCAL_BASE_URL
  • Run npm run rag:rebuild:zero
• If Ollama/model remains unavailable:
  • Continue in lexical mode (non-blocking fallback).
  • Mark risk explicitly in tracker/snapshot (rag degraded, confidence reduced).
  • Keep MCP/LSP diagnostics active.
• Long rebuild:
  • Keep process running; full profile can take minutes.
  • If markdown volume is delaying index, run with --docs-mode exclude and process docs in a second pass.

Command Reference

• npm run daemon:start
• npm run daemon:status
• npm run mcp:diagnose
• npm run rag:rebuild:zero
• npm run rag:rebuild:zero -- --docs-mode exclude
• npm run rag:health -- --json
• npm run rag:index -- --profile full --docs-mode only

Done Criteria

• PM2 shows agente-gpt, dashboard-web, chrome-proxy online.
• http://localhost:3008/api/mcp responds.
• mcp:diagnose passes.
• RAG rebuild completes with exit code 0, or fallback lexical explicitly registrado com risco.