vedun
Apache-2.0 · local-first · v0.1.0 (early release)

A local knowledge base for standards, specifications and other technical documents.

Ask in plain language, get the exact clause, its edition and its verbatim text, with a ready-to-paste citation.

Install on macOS / Linux
curl -fsSL https://raw.githubusercontent.com/numikel/vedun/main/scripts/install.sh | bash
Windows, or installing with an AI agent →
vedun dashboard: a clause found by its number, with a ready-to-paste citation and the verbatim text
01

Why vedun

Who it is for

vedun is built for people who cite documents for a living, and for the AI agents that help them.

  • Engineers
  • Auditors
  • Health and safety and compliance specialists
  • AI agents working alongside them, through the MCP server

It never invents a clause number

A citation is only useful if it can be checked. vedun answers from the documents you gave it, and nothing else.

  • Every answer points to a real unit of a real edition
  • You get the clause, its edition and its verbatim text
  • The quote can be checked against the PDF before you send the email or the report
02

How it works

vedun dashboard overview: quick search, Sync button and counts of indexed documents, clauses, tables and figures
  1. 1

    Drop your PDFs into the pdf folder

    Name each file <doc_id>_<YYYYMM>.pdf, for example ISO-45001_201803.pdf. The documents stay on your computer.

  2. 2

    Sync

    Docling converts the PDFs locally. vedun then builds a full-text index (SQLite FTS5) and a vector index (sqlite-vec).

  3. 3

    Ask

    Ask in the dashboard, or let your AI agent ask through the MCP server.

03

Features

Exact lookup

Go straight to a clause, table or figure by its number, for example 4-12.

Search by meaning, in any language

Describe what you are looking for in your own words. You can ask in Polish about documents written in English.

Full-text search

Find exact phrases and numbers with SQLite FTS5.

verify_quote

Checks that a sentence really is in the clause you cite, and in the text layer of the PDF.

Edition changes

A report of removed, new, changed and probably renumbered clauses between editions. check-citations finds the references in your notes that no longer match.

Dashboard, REST API and MCP server

A dashboard in English and Polish, a REST API, and an MCP server with 10 read-only tools: kb_lookup, kb_search, kb_fulltext, kb_verify_quote and more.

Local-first

Documents, database and files stay on your computer. Only embeddings (OpenRouter) and the optional reranker go online, and only with your own key.

04

Installation

You need Docker Desktop (Windows, macOS), Docker Engine or Podman (Linux), about 4 GB of disk and 8 GB of RAM (16 GB for long PDFs).

Install script
curl -fsSL https://raw.githubusercontent.com/numikel/vedun/main/scripts/install.sh | bash
Full instructions in the README. Container image: ghcr.io/numikel/vedun
05

FAQ

Only the text sent for embeddings, and only when you provide a key. Without a key, exact lookup and full-text search still work. Documents, database and files stay on your computer.

vedun is free and open source. Embeddings are paid per token to the provider, once per document. A daily budget limits the spend; the default is 3 USD.

PDFs with a text layer. Name each file <doc_id>_<YYYYMM>.pdf. The pattern of clause numbers is set in a JSON profile.

It listens on 127.0.0.1 only and fails closed. Access needs an API key and is rate limited. The container runs as non-root with a read-only root file system.

Apache-2.0.