convez.ai
MCP ServerBusca & PesquisaIntermediário
gvtret

doc-rag MCP Server

gvtret/doc-rag-mcp-server

RAG local para documentação técnica via MCP

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP para RAG local e offline-first sobre documentos de engenharia (PDF, DOCX, DOC, MD, TXT). Integra-se nativamente com Cursor e Claude, mantendo todos os dados na máquina sem conexão externa. Inclui UI web para upload, ingestão e gerenciamento.

Instalação

$npx convez-ai gvtret/doc-rag-mcp-server

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
1
Saúde
71%
Último commit
há 2 dias
Versão
v2.2.2
Issues abertas
0
Contribuidores
2

Casos de uso

  • Indexar e pesquisar bases de documentação técnica privada
  • Integrar buscas semânticas em Cursor/Claude mantendo dados offline
  • Criar assistentes IA sobre manuais e specs de engenharia

Veredito

Solução madura e bem arquitetada para RAG local com excelente suporte a múltiplos formatos e integração MCP nativa. Ideal para equipes que precisam de IA sobre documentação técnica sem enviar dados para a nuvem.

README

A tradução para o português deste README ainda está na fila.

doc-rag — local RAG over your documents

tests lint build license: MIT

doc-rag is a local, offline-first knowledge base for engineering documentation:

PDF / DOCX / DOC / MD / TXT  →  Markdown  →  Chunks  →  FAISS  →  MCP / Cursor / Web UI
  • Offline after first model download. No data leaves your machine.
  • Cursor / Claude MCP integration via Streamable HTTP on a single endpoint.
  • Web UI for upload, ingest, delete, and live status — no terminal required.
  • Graceful degradation: if FAISS isn't ready, lexical search keeps working and the user is told.

Requirements

  • Linux or WSL2 (Python ≥ 3.10)
  • ~2 GB RAM for embeddings; ~1 GB extra for Docling models on first parse
  • antiword is optional (only needed for legacy .doc files)
  • OCR for scanned PDFs is built into Docling (RapidOCR); no separate Tesseract install required
  • Node ≥ 20 (optional, v2.2+ — only needed to build the new Svelte /ui-next/ page; the legacy inline /ui works without Node)

Quickstart

# uv is the official installer since v2.1; the bootstrap script installs it if missing.
git clone https://github.com/gvtret/doc-rag-mcp-server
cd doc-rag
bash scripts/bootstrap.sh        # installs uv, runs `uv sync --frozen`, creates .venv
cp YOUR_FILES.pdf sources/incoming/
uv run doc-rag ingest
bash scripts/run_mcp_http.sh     # MCP/UI on http://127.0.0.1:3333

Then open http://127.0.0.1:3333/ui or point Cursor at http://127.0.0.1:3333/mcp.

Documentation

GuideWhat's inside
docs/install.mdvenv + torch (CPU/GPU), system packages, OCR, config reference
docs/cli.mddoc-rag ingest / rebuild / delete / wipe / clean-orphans / clear-incoming
docs/mcp.mdCursor / Claude integration, Streamable HTTP + SSE, auth, rate limit
docs/ui.mdWeb UI: upload, dedup, ingest, delete, danger zone, degraded-mode banner
docs/deploy.mdDocker Compose, native systemd, deploy archive, remote MCP
docs/troubleshooting.mdTorch/CUDA, MCP not visible, FAISS rebuild, PEP 668, OCR
docs/roadmap.mdVersioning policy and the path to public v1.x.y

See CHANGELOG.md for notable changes between releases.

Project layout

doc-rag/
├── sources/
│   ├── incoming/      # drop new documents here
│   └── archived/      # processed files (moved automatically after ingest)
├── build/             # generated: docs_md/, chunks_jsonl/, embeddings/, index/, manifest.json
├── config/config.yaml # main config
├── src/doc_rag/       # Python package
├── scripts/           # bootstrap, run_mcp_http, install_server_native, ...
├── docker/            # Dockerfile
├── systemd/           # service unit template
├── .github/workflows/ # CI (tests, lint, Docker build)
└── docs/              # see the table above

Philosophy

Offline-first · reproducible · vendor-independent · long-term maintainable. Designed for standards, specs, manuals, and research documents.

Talks and articles

License

doc-rag is licensed under the MIT License — see LICENSE.

Third-party dependency licenses are summarised in NOTICE. Releases before v2.0.0 were AGPL-3.0-or-later (because the default PDF backend was PyMuPDF); v2.0 switched to Docling, which is MIT, and relicensed the project to match.

Contributing

Issues and pull requests are welcome. Please read CONTRIBUTING.md for development setup, test instructions, commit conventions, and the SemVer policy for the public surface (docs/roadmap.md § 1).

To report a security issue, see SECURITY.md. Please do not open a public issue for vulnerabilities.

Ver o original no GitHub

Bom para

Relacionados