rag-mcp-server Skill Skill

rag-mcp-server Skill

Expose a local BM25 knowledge-base and session index to any MCP client — Claude Code, Claude Desktop, or any gptme instance — so queries like "how does virtual memory work?" resolve to book chapters inline, and "what did we discuss about caching last week?" resolve to past sessions.

Overview

rag-mcp-server is part of the rag package in your gptme agent workspace. It exposes two search surfaces as MCP tools:

| Tool | What it searches | |------|-----------------| | search_wisdom | BM25 index over CC-licensed CS reference books | | list_wisdom_sources | List of book slugs currently indexed | | search_sessions | Journal, gptme logs, and Claude Code trajectory index |

Default book set (7): SICP, OSTEP, Pro Git, Think Python, MML Book, RL Intro, Eloquent JavaScript. BM25 precision@3: 0.87 on a curated benchmark set.

Setup

1. Install the package

# From gptme-contrib workspace root
uv sync --all-packages

# Verify the entry point is available
rag-mcp-server --help

2. Build the book (wisdom) index

# Download and index CC-licensed books in one command:
bob-search index-books --preset cs-fundamentals

# Preview without downloading:
bob-search index-books --preset cs-fundamentals --dry-run

# List available presets:
bob-search index-books --list-presets

# For books requiring manual download (follow --dry-run instructions, then):
#   gptme-wisdom ingest --source sicp --file ~/books/sicp.md
# Supported manual-ingest slugs: sicp, ostep, pro-git, thinkpython, mml-book,
#   rl-intro, eloquentjs

Default wisdom DB: ~/.local/share/gptme/wisdom.db

3. Build the session index (optional)

bob-search index    # indexes journal/, gptme logs, Claude Code trajectories

Default session DB: ~/.local/share/bob/session-index.db

4. Run a quick smoke test

# Verify the wisdom index is queryable (should return OSTEP chunks):
bob-search wisdom "virtual memory" --source ostep --top-k 2

# Verify the session index is queryable (if indexed):
bob-search sessions "caching" --limit 2

# Verify the MCP server starts and responds to JSON-RPC:
echo '{"jsonrpc":"2.0","method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"0.1"}},"id":1}' | \
  rag-mcp-server | head -1

MCP Configuration

Claude Code

Add to .claude/settings.json in your project (or ~/.claude/settings.json for global use):

{
  "mcpServers": {
    "wisdom-rag": {
      "command": "rag-mcp-server",
      "type": "stdio"
    }
  }
}

Or add it from the terminal:

claude mcp add wisdom-rag rag-mcp-server

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "wisdom-rag": {
      "command": "rag-mcp-server",
      "args": []
    }
  }
}

Custom DB paths (non-default install)

{
  "mcpServers": {
    "wisdom-rag": {
      "command": "rag-mcp-server",
      "args": [
        "--wisdom-db", "/path/to/wisdom.db",
        "--sessions-db", "/path/to/sessions.db"
      ]
    }
  }
}

Tools Reference

These MCP tools become available to the agent once the server is configured. Call them when a question would benefit from grounding in CS fundamentals or past session history.

`search_wisdom`

search_wisdom(query: str, source?: str, top_k?: int = 5) -> list[Chunk]

Searches the indexed book corpus for passages relevant to the query. Each result includes: source (book slug), title, chapter, content (excerpt), score.

When to call: the user asks "how does X work?" or you need to ground a technical explanation in authoritative CS references.

Example queries that work well:

"how does virtual memory work" → OSTEP Chapter 13
"git rebase vs merge" → Pro Git
"reinforcement learning q-learning" → RL Intro
"higher-order functions in JavaScript" → Eloquent JS
"backpropagation gradient descent" → MML Book

`list_wisdom_sources`

list_wisdom_sources() -> list[str]

Returns book slugs currently indexed (e.g. ["ostep", "sicp", "pro-git", ...]).

When to call: before a search_wisdom call if you're unsure which sources are available, or to verify the index is populated.

`search_sessions`

search_sessions(query: str, source?: str, limit?: int = 5) -> list[Session]

Searches past sessions (journal, gptme logs, Claude Code trajectories). Each result includes: source, date, title, summary, path, score.

source can be journal, gptme, or claude_code.

When to call: the user asks "what did we discuss about X?" or you need context from a previous session to inform current work.

Example queries:

"RAG retrieval implementation" → sessions where RAG was discussed
"bug in websocket reconnect" → debugging sessions from the past
"how did we approach the authentication refactor" → design sessions

Usage Examples

Once configured, Claude calls these tools automatically during conversation:

User: How does copy-on-write work in operating systems?
Claude: [calls search_wisdom("copy-on-write operating system", source="ostep")]
        → OSTEP §21: "…when a child forks, pages are shared read-only;
           a write triggers a private copy…"

User: What did we implement for the session recording last week?
Claude: [calls search_sessions("session recording implementation", limit=3)]
        → Journal 2026-06-28: "Implemented incremental re-index…"

Troubleshooting

rag-mcp-server not found: run uv sync --all-packages from the gptme-contrib root, then verify with which rag-mcp-server.

No wisdom sources indexed: run bob-search index-books --preset cs-fundamentals to build the initial wisdom DB.

search_sessions returns empty: run bob-search index to build the session index. Journal-only installs only need the journal path to exist.

MCP server unreachable in Claude Code: confirm the rag-mcp-server command is on PATH in the shell Claude Code uses (try claude mcp list to verify the server is registered and claude mcp get wisdom-rag to see its status).

Package source: packages/rag/ in your gptme agent workspace
Full docs: packages/rag/README.md in your gptme agent workspace
MCP specification: modelcontextprotocol.io
gptme-wisdom ingest CLI (upstream): gptme-wisdom --help

Agent Skills: rag-mcp-server Skill

Install this agent skill to your local

Skill Files