Enterprise Knowledge MCP Server

MCP server that exposes semantic search + RAG over enterprise documents as standardized tools, so any MCP client (Claude Desktop, etc.) gets read-only, citation-backed access without touching source systems directly.

Connector #1 (local PDF/DOCX/XLSX) is implemented. Embeddings run locally via sentence-transformers — no document text leaves the machine.

Architecture

connector.list_documents()          # SourceDocument objects
        v
ingestion/extractor.py              # per-mime text extraction -> ExtractedSection
        v
ingestion/chunker.py                # sentence-packed chunks w/ overlap -> Chunk
        v
ingestion/embedder.py               # sentence-transformers -> vectors
        v
storage/vector_store.py             # Chroma upsert/query (metadata filtering)
        v
rag/retriever.py (KnowledgeService)  # orchestrates the above; search() + build_rag_context()
        v
server.py                            # FastMCP tools wrapping KnowledgeService

Adding a new source (SharePoint, Confluence, Notion, Google Drive, Slack) means writing one class implementing connectors/base.py::BaseConnector (list_documents, get_document). Nothing downstream changes.

Setup

pip install -r requirements.txt

Drop files into data/raw/ (or point at another directory), then from src/:

from connectors.local_files import LocalFileConnector
from rag.retriever import KnowledgeService

svc = KnowledgeService()
svc.ingest_connector(LocalFileConnector(root_dir="../data/raw"))

Or just run the ingest tool through the MCP server itself once connected.

Running the server

python src/server.py

Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "enterprise-knowledge": {
      "command": "python",
      "args": ["/absolute/path/to/enterprise-knowledge-mcp/src/server.py"]
    }
  }
}

Tools exposed

Tool	Purpose
ingest_local_directory(directory?)	Scan + index PDF/DOCX/XLSX from a directory
search_knowledge(query, top_k?, source_type?, department?, author?)	Semantic search with metadata filters
get_rag_context(query, top_k?, source_type?, department?)	Retrieved chunks + citation markers, ready for an LLM prompt
list_indexed_documents()	All indexed docs + chunk counts
index_stats()	Total indexed chunk count

Metadata schema (per chunk)

doc_id, title, source_type, path_or_uri, section_label, author, department, tags, modified_at, ingested_at — section_label is page number (PDF), heading/table (DOCX), or sheet name (XLSX), so citations point to a specific location, not just a document title.

Config

Environment variables (see src/config.py): EKM_RAW_DIR, EKM_CHROMA_DIR, EKM_COLLECTION, EKM_CHUNK_SIZE, EKM_CHUNK_OVERLAP, EKM_EMBEDDING_MODEL, EKM_TOP_K.

Roadmap (modular connectors, in priority order)

1. Local files — done.
2. SharePoint — Microsoft Graph API (/sites/{id}/drive/items), delta queries for incremental sync, app-only auth via client credentials.
3. Confluence — REST API v2 (/wiki/api/v2/pages), CQL search for incremental sync, API token auth.
4. Notion — Notion API (/v1/search, block children pagination), integration token auth.
5. Google Drive — Drive API v3, service account or OAuth, export Google Docs/Sheets as text/csv before extraction.
6. Slack — conversations.history + file downloads, bot token auth, respect channel membership for access control.

Security notes for production

• Never embed credentials in connector code — use env vars / a secrets manager per connector.
• Enforce source-level ACLs at query time (filter where clauses by the requesting user's permitted departments/channels), not just at ingest time.
• Log every search_knowledge / get_rag_context call (query + returned doc_ids) for audit trails.
• Rate-limit and sandbox connector network calls; SharePoint/Confluence/ Notion/Drive/Slack connectors will need outbound HTTPS to their respective APIs only.