MCP tools

rag-hub-mcp exposes 10 tools over the Model Context Protocol. Any MCP-compatible agent (opencode, Claude Code, Cline…) can call them.

Tool	Description
rag_list_kbs	List knowledge bases with stats
rag_list_documents	List documents in a KB
rag_search	Search KBs (kb optional — omitting searches all). Write a self-contained query with full context
rag_add_document	Add a text document (creates folder if needed)
rag_read	Retrieve the full extracted content of a document (KB + path)
rag_delete_document	Delete a document
rag_delete_kb	Delete an entire KB
rag_reindex	Scan for changes (force: true rebuilds from scratch) — enqueues jobs, returns counts
rag_status	Index overview (KBs, documents, chunks) + queue stats
rag_jobs	Indexing queue status (pending/processing/failed) and recent failures

Example calls

{
  "name": "rag_list_kbs",
  "arguments": {}
}
// → 3 KBs, 42 documents, 128 chunks

{
  "name": "rag_search",
  "arguments": { "query": "how to set up SSO", "kb": "infra", "top_k": 10 }
}
// → ranked chunks with source citations

{
  "name": "rag_add_document",
  "arguments": {
    "kb": "dev",
    "path": "notes/setup.md",
    "content": "# Setup\n…"
  }
}
// → added + queued for indexing

rag_search

• query (required): natural-language search query. Write it self-contained with full context (e.g. "API v2 rate limits", not "the other version") — the server is stateless, so follow-up references from the conversation are resolved client-side before the call.
• kb (optional): restrict to one knowledge base.
• top_k (optional, default 10): number of results.

Each result's content is truncated to SEARCH_SNIPPET_MAX_CHARS (default 1000, see Configuration). The cap is independent from the indexing-time CHUNK_MAX_CHARS, so a retrieved chunk may be longer than its snippet. Use rag_read to get the full content of a document when the relevant part sits past the cap.

rag_add_document

• kb (required): knowledge base name; the folder is created if needed.
• path (required): relative path in the KB (e.g. notes/architecture.md).
• content (required): text or Markdown content.

The document is queued for indexing and processed asynchronously by the worker (extract → chunk → embed → insert). Use rag_jobs to check progress.

rag_reindex

• force (optional, default false): trigger a full rebuild instead of an incremental scan. When true, all chunks and files are purged from the index (KB folders are kept), then every document is re-queued for indexing. Use it after enabling contextual chunking or changing an indexing option, so previously indexed documents pick up the new embeddings. Without force (or when the argument is omitted) it is a plain incremental scan — unchanged files are skipped, new/modified files are enqueued.

The scan returns added/modified/deleted/skipped/excluded counts. Actual indexing happens asynchronously in the worker — results appear once the worker has processed the enqueued jobs.

rag_jobs

Queue status without arguments.

Queue: 2 pending, 1 processing, 0 failed

Failed jobs are parked after INDEXER_RETRY_MAX attempts. Retry them manually via the REST API (POST /admin/jobs/:id/retry).