UNHRDB is a research environment for finding and citing what UN human rights mechanisms actually wrote. It treats UN documents as structured, paragraph-level research material while keeping every result connected to its original document symbol and context.
- Live application: https://lszoszk.github.io/generalcomments/
- General Comments dataset: https://huggingface.co/datasets/lszoszk/treaty-bodies-general-comments
- MCP integration: https://github.com/lszoszk/mcp-unhrdb
- Claude skills: a plugin with the connector and three research skills — install; skills for the recommendations to States come with UHRI+
- Methodology: METHODOLOGY.md
- Document status audit: STATUS_AUDIT.md
The October 2026 public build contains 326,448 searchable paragraphs across 6,152 documents:
| Collection | Coverage |
|---|---|
| Treaty Body General Comments and Recommendations | 188 documents |
| Individual-communication jurisprudence | 4,333 cases from 8 mechanisms |
| Thematic Special Procedures reports | 1,631 reports from 46 mandates |
| Recommendations to States (UHRI, live) | 267,942 records from 70 mechanisms to 199 States, 2006–2026 |
The Recommendations scope is not part of the static corpus: concluding
observations, Universal Periodic Review and country-visit recommendations are
queried live from the UHRI dataset service this project runs on its server (a
mirror of the OHCHR Universal Human Rights Index export); only the filter
vocabularies ship with the site (docs/rec/vocab.json).
Coverage changes as the source collections are synchronised. The application publishes build dates, collection-level provenance, and known gaps alongside the data.
- Searches legal text at paragraph level with exact word forms and phrases,
implicit/explicit
AND,OR, grouped exclusions (A NOT BorA -B), explicit prefix wildcards, and filters. A positive term is required for exclusions so every query stays on the indexed path; negative-only searches are rejected with an explanation. - Keeps General Comments, jurisprudence, and Special Procedures as distinct source collections rather than blending their authority or provenance.
- Checks the UN citations in a pasted text (
#check): symbols, "General Comment No. N", "Communication No. N/YYYY" and case names are resolved against the catalogue, superseded texts and invented ones are flagged, pinpointed paragraphs are verified and quotations compared word by word with the source, all in the browser. - Opens results in a source-first document reader with paragraph identifiers, context, bookmarks, notes, and citation export.
- Reconstructs and audits document structure, OCR provenance, footnotes, and links to authoritative UN records.
- Provides an experimental extractive Ask workflow that returns verbatim Committee passages rather than generated legal answers.
- Exposes General Comments and UHRI recommendations to MCP-capable clients via
the companion
mcp-unhrdbserver. - Ships Claude skills
for researching UN human rights law with verbatim, paragraph-cited sources,
checking UN citations in a text and analysing treaty-body case law (skills for
recommendations to States are published with UHRI+). Install in Claude Code with
/plugin marketplace add lszoszk/mcp-unhrdband/plugin install unhrdb@mcp-unhrdb.
The published application is static and can be served directly from docs/:
git clone https://github.com/lszoszk/generalcomments.git
cd generalcomments
python3 -m http.server 8000 --directory docsThen open http://127.0.0.1:8000.
The General Comments and Recommendations corpus is distributed as Parquet on
Hugging Face.
This repository holds the published site. The preparation, ingestion,
validation, and audit code that produces the build is kept in a separate,
private repository; researchers who need it can contact the authors.
PUBLIC_FILES lists what this repository may contain, and a
check on every push enforces it. See
METHODOLOGY.md and the public coverage information in the
application for limitations and provenance.
The Hugging Face package covers General Comments and Recommendations. The jurisprudence and Special Procedures collections are not currently distributed as part of that dataset package.
For the evolving software, use the concept DOI:
Szoszkiewicz, Ł., & Kowalska, Z. (2026). UNHRDB — UN Human Rights Database (Version 2.2.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.10495719
GitHub also exposes machine-readable citation metadata from
CITATION.cff. Cite a fixed Zenodo version when an exact
software snapshot is required for reproducibility.
When relying on a legal proposition or quoting a paragraph, cite the original
UN document symbol and printed paragraph number, for example CRC/C/GC/25 ¶12.
The database citation supplements rather than replaces the primary legal
source.
docs/citations/graph.json lists every reference one document makes to
another (UN symbols, "general comment No. N", "communication No. N/YYYY",
in bodies and footnotes), with the citing paragraph ids; graph.csv holds the
same edges as CSV, and index.json the per-document adjacency the reader uses
for "Cited by" / "Cites". The files are rebuilt with every corpus change and
registered in docs/manifest.json.
This repository contains distinct works under distinct terms:
- Software: PolyForm Noncommercial License 1.0.0; research, education, non-profit, and personal use are permitted, while commercial use requires a separate licence. See LICENSE.
- Curated dataset and annotations: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International, see LICENSE-DATA.
- Underlying UN documents: remain subject to United Nations content terms and are not re-licensed by this project.
See NOTICE for the detailed boundary between software, curation, and source documents.
- Łukasz Szoszkiewicz, Adam Mickiewicz University, ORCID 0000-0001-6671-2893
- Zuzanna Kowalska, Adam Mickiewicz University
Issues, missing documents, and corrections can be reported through the GitHub issue tracker. General contact: l.szoszkiewicz@amu.edu.pl.