Citewell

Citation-grounded document memory for AI assistants. Offline · MCP-native · Turkish-ready.

Citewell is a headless retrieval server. It ingests your PDFs, Markdown, text, HTML and Word files into PostgreSQL + pgvector and serves cited passages — file, page, chunk id and a verbatim quote — to AI assistants over the Model Context Protocol, a CLI and a REST API.

It never calls an LLM and needs no API key: your assistant (Claude, Cursor, …) writes the answer, Citewell makes sure it can point at the source.

• Getting started — docker compose up, ingest, search.
• Connect an assistant — Claude Code, Claude Desktop, Cursor.
• Architecture — ingest and query pipelines, data model.
• Evaluation — Recall@k, MRR, page hit rate, latency.

What a search returns

{
  "ref": "[1]",
  "citation": "calisan-el-kitabi.pdf, s. 4",
  "source": "calisan-el-kitabi.pdf",
  "page": 4,
  "chunk_id": "c9a6068735629cb0647e",
  "score": 0.0328,
  "quote": "İş seyahatlerinde konaklama için gecelik üst sınır İstanbul, Ankara ve İzmir'de 3.500 TL, diğer şehirlerde 2.500 TL'dir.",
  "text": "4. Masraf ve Seyahat\nİş seyahatlerinde konaklama için …"
}