SELECT A DOCUMENT
Pick any document from the rail on the left to read its full text. Bookmarks, notes and citations stay attached to specific paragraphs.
Reading the paragraph-level corpus from your archive.
MOBILE MODE Simplified search for this device — exact words only (variants like “child/children” and operators like AND/OR aren’t matched). Open on a desktop for full search.
Pick a document on the left to read it in full. Notes and bookmarks appear on the right.
Pick any document from the rail on the left to read its full text. Bookmarks, notes and citations stay attached to specific paragraphs.
Bookmarks, notes, pinned comparisons and saved searches live in your browser's localStorage. Nothing leaves your device. Clear browser data and they're gone.
Paste a paragraph, a footnote block or a whole section. Every UN document symbol, “General Comment No. N”, “Communication No. N/YYYY” and “X v. State” is resolved against this database; a pinpointed paragraph is checked to exist, and a quotation is compared word by word with the source paragraph. Superseded texts are flagged, invented ones are marked in red. Your text stays in this browser: only the identifiers of cited documents are fetched.
A paragraph-level companion to Treaty Body General Comments and Special Procedures, with selected jurisprudence.
UN Human Rights Database is a paragraph-level search interface for UN human rights interpretation. Its authoritative core is Treaty Body General Comments. It now also indexes the complete set of thematic Special Procedures annual reports — all 46 mandates — and a near-complete set of CCPR / CEDAW / CRPD individual-complaint jurisprudence — about 98% of the OHCHR JURIS catalogue. Special Procedures reports are soft law (independent expert opinion, not binding treaty-body interpretation); the jurisprudence collection is near-complete but not yet exhaustive — a small non-English / image-only-scan remainder is pending. Secondary collections are loaded lazily on demand. A fourth scope, Recommendations, searches what the same mechanisms said to one State — concluding observations, Universal Periodic Review and country-visit recommendations — served live from the OHCHR Universal Human Rights Index (see below); it is not part of the paragraph counts here. Across the current published build, 322,619 paragraphs across 6,111 documents are searchable, filterable and citable down to their number.
This is an independent academic project, not affiliated
with or endorsed by the OHCHR or any UN entity. Citations of individual
paragraphs should reference the original UN document signature (e.g.
CRC/C/GC/25 ¶12), not this database.
Dataset snapshot: last synchronised with OHCHR Treaty Body Database on 3 September 2026. 6,111 documents · 322,619 paragraphs total — 187 General Comments (all 10 treaty bodies), 1,591 Special Procedures reports (all 46 thematic mandates), 4,333 jurisprudence cases (8 bodies: CCPR, CAT, CEDAW, CESCR, CRC, CRPD, CERD, CED). The most recent General Comment is CEDAW/C/GC/30/Add.1 (Women, Peace and Security addendum, 18 February 2026); the most recent Special Procedures report is A/HRC/63/36 (SR Contemporary Slavery, Contemporary forms of slavery and trafficking in persons, August 2026). Weekly link checks run on GitHub Actions. See change log for previous snapshots.
This is the completed and most authoritative part of the app: 187 General Comments issued by the ten UN human rights treaty bodies (CEDAW, CCPR, CERD, CESCR, CRC, CRPD, CMW, CAT, CAT-OP, CED). The collection is exhaustive as of build date, sourced primarily from tbinternet.ohchr.org, and annotated for concerned groups such as children, women, migrants and persons with disabilities.
4,333 individual-communication cases ·
152,731 paragraphs from all 8 OP / individual-complaint
mechanisms — CCPR (2,861), CAT (835), CEDAW (172), CESCR (155),
CRC (150), CRPD (95), CERD (58), CED (7). Published as separate
shards under docs/jur/ so the main General
Comments interface stays lightweight; the shard for a case is
fetched lazily when the user opens it. CCPR and CAT include
OCR-recovered scanned legacy cases (some from the early 1990s with
no embedded text layer); the reader-pane displays a provenance
banner on every OCR-recovered document with a one-click
report-issue link to the GitHub tracker.
Sources & cross-checks. Primary source: the OHCHR bulk jurisprudence dump (catalog + raw English DOCX/PDF). For partial cross-checking and methodological inspiration in selected areas — particularly metadata normalisation and substantive-article extraction for CCPR cases — we also consulted external curated databases such as the CCPR Centre's decisions digest and the University of Minnesota Human Rights Library.
We match
3,950 of the 4,015 unique case symbols
(98.4 %) listed in the
OHCHR JURIS authority catalogue.
The 65 catalogue entries we don't yet cover fall into a
few well-understood buckets — when we audited the missing
list, 43 had no PDF available in any language
on OHCHR's own download surface (empty downloads
dict on the catalogue entry), so they're not recoverable from
OHCHR; the rest split across:
Per-treaty coverage: CRPD 100 % · CED 100 % · CRC 99.3 % · CAT 98.8 % · CCPR 98.4 % · CESCR 98.4 % · CEDAW 98.0 % · CERD 75.4 %.
The full list of 65 missing decisions — missing-cases.csv on GitHub — has the OHCHR JURIS catalogue URL for every entry plus the source-recovery notes (methodology & sources tried). For research help on a specific case, contact l.szoszkiewicz@amu.edu.pl.
1,591 reports covering all 46 thematic mandates of the UN Special Procedures — 33 Special Rapporteurs, 6 Independent Experts and 7 Working Groups. Each mandate's full run of thematic annual reports to the Human Rights Council, the General Assembly and the former Commission on Human Rights is split paragraph by paragraph and labelled for concerned groups, the same way as General Comments. Special Procedures reports are soft law: the independent opinion of a mandate-holder, persuasive but not the binding interpretation a treaty body issues in a General Comment. Footnotes: 1,415 of the 1,591 reports carry their footnotes as clickable references — 74,401 notes anchored at the reference, recovered on 6 September 2026 from the Word files UN Documents serves; 79 reports have no Word file there and 75 older ones have a Word file without notes, so a bare digit in those texts is a reference whose note could not be recovered.
What is in and out of scope. The corpus is the thematic output only. By design it excludes country-visit reports, communications (joint letters and urgent appeals) and addenda / corrigenda — these report on situations rather than develop doctrine. A small number of older reports are not yet included: a handful are unfetchable through the UN document API (an internal-redirect glitch, retried periodically) and a few are image-only scans with no text layer (pending OCR). Reports are added and maintained through a deterministic ingestion pipeline.
The Recommendations scope adds the layer every research guide lists first: what the treaty bodies said to a particular State in their concluding observations, what the Human Rights Council recommended in the Universal Periodic Review, and what Special Procedures recommended after a country visit. The records are the OHCHR Universal Human Rights Index annotations — 267,942 records from 70 mechanisms addressed to 199 States, 2006 to 2026 — each tagged by OHCHR with themes, affected persons and Sustainable Development Goals. Those tags are the filters; a “Recommendations only” switch drops the concerns and observations that precede them. Records are ordered by publication date, newest first; the service has no relevance ranking.
How it is served. Unlike the three collections above, the records are not part of this site’s static corpus. Every search asks the UHRI dataset service that this project runs on its server (a mirror of OHCHR’s export, refreshed from OHCHR; the newest record in the current mirror was published on 20 May 2026), so the scope needs a connection and is absent from the offline copy, the Ask tab and the citation graph. Each record links to its document on UN Documents and to its page in UHRI; the dossier lists every record of the same document. Citation formats treat a record as a paragraph of the concluding observations, UPR Working Group report or visit report it comes from. The filter vocabularies are rebuilt by build_rec_vocab.py.
Each document is downloaded as PDF, converted to Markdown, then to
structured JSON paragraph-by-paragraph. Paragraphs are annotated with
concerned-group labels from a fixed 19-category taxonomy
(children, women/girls, persons with disabilities, persons deprived of
their liberty, indigenous peoples, LGBTI+, and so on). The corpus is
rebuilt deterministically by
build_corpus.py;
local fallback search runs in the browser via FlexSearch and verifies
exact word forms; a trailing wildcard such as child*
explicitly includes variants. The index is cached in IndexedDB after
first load.
One caveat on references: structured cited-article metadata is
currently extracted for jurisprudence only (about 72 % of cases). In
General Comments and Special Procedures, treaty-article citations stay
fully searchable inside the footnote and body text but are not yet pulled
into a separate field.
The Ask tab is an experimental retrieval layer over the General Comments corpus. It returns verbatim paragraphs from the Committees — no paraphrase, no AI-generated answer text. The LLM only operates on the query side (rewriting the question into doctrinal language) and as a final-stage ranker (deciding which of 50 retrieved paragraphs most directly answer the question). This is an extractive RAG — fundamentally different from the chatbot style that synthesises answers and may hallucinate citations.
30-question commentary-anchored benchmark (Saul on ICESCR, Joseph & Castan on ICCPR; 16/30 carry paragraph-level expected IDs).
Cross-treaty references inside GCs (a CERD/CRC/CEDAW GC mentioning
"the Covenant" or "article 4(2)" meaning ICCPR) are handled by a
defensive fallback: "article N of that Protocol" is left
as plain text rather than wrong-linked. Full disambiguation needs
the corpus pre-tagged with explicit cited_articles
metadata — deferred to a later pipeline. Some lay queries surface
adjacent-but-not-canonical paragraphs from the right document —
usable, sometimes confusing. Per-IP rate limit (15 q/min) caps
token costs from automated traffic; researchers running heavy
iteration may briefly hit it.
Treaty-text augmentation, the LLM-judge stage, and the rerank cross-encoder all operate on actual paragraph text, not on summaries. Every visible citation is traceable: signature + GC number + paragraph number → click Open dossier for full context. The Ask tab augments — it does not replace — paragraph-level Search, which remains the source of truth for exact wording or Boolean queries.
170 older CCPR cases (mostly 1980s–90s scanned PDFs) come into
the corpus through Tesseract via a quality-first sidecar pipeline.
5,359 paragraphs across these cases are
OCR-recovered; the remaining 106,688 jurisprudence paragraphs
come from native text-layer PDFs or DOCX. Published paragraphs
keep official numbering where possible; individual opinions are
namespaced as OP1-,
OP2-, and real case appendices as
A1-. Generated IDs and OCR provenance
stay exposed in the export rather than silently merging into
official text.
Cleanup pipeline (May 2026): a
five-stage post-OCR sweep brought 1,388 fixes across
~422 paragraphs. Stages: (1) mechanical fn-marker and
section-header repairs (436); (2) Tesseract tsv
word-row leak strip — 575 cases where coordinate metadata had
bled into paragraph text (e.g.
August 19825 1 6 1 4 5 1049 1995 130 42 49.939957 from
→ August 1982 from); (3) hand-vetted
mangled-word audit on 229 prefilter-flagged ¶ — Spanish-name
accent recovery (Sanjuán, Gaitán, Velásquez Rodríguez),
sentence-start I mis-OCR'd as 1, and well-known
case-name fixes (47); (4) ~80 letter-substitution rules covering
systematic Tesseract shape confusions
(I↔l, u↔n, c↔e, h↔b,
m↔in, g↔s) — 289 fixes; (5) end-to-end spot-read
of 10 heavily-fixed paragraphs catching one regression
(beae. iciary wrongly normalised to
judiciary; correct reading was beneficiary).
Every rule is verified in context; nothing was applied where the
source word could be a real English/Spanish/French token.
Residual errors: non-systematic
single-character glitches remain (e.g. cn
for on, i:to for into).
The reader pane carries an OCR provenance banner
on every scanned-PDF case with a one-click
report-issue link.
If you spot one in your research, please file it — the cleanup
pipeline is open to community contributions and audit.
Labels applied manually before this pass are preserved verbatim; the automated pipeline only fills paragraphs left unlabelled. Pipeline labels come from human-readable regular expressions matched against paragraph text — every match is deterministic and reproducible. After six iterations, the corpus reaches 74 % coverage on General Comments (5,283 of 7,105 paragraphs labelled) and 37 % on Special Procedures (about 6,990 of 18,751 paragraphs labelled). The 26 % of GC paragraphs that remain unlabelled are mostly procedural framework text or discuss groups outside the taxonomy (older persons, persons of African descent, religious minorities).
Patterns target unambiguous markers — full terms and fixed phrases
such as indigent,
free, prior and informed consent,
corporal punishment — rather than generic
synonyms. During quality review we removed several patterns that
produced false positives: social security
and social protection (wrongly mapped
to Poverty), forced displacement (moved
from Refugees to Armed conflict), marginalized
(removed from Poverty in SP because it overwhelmingly described
religious minorities in freedom-of-religion reports), and
alien (too ambiguous for Non-citizens).
81 false-positive Poverty labels were cleaned out in this pass;
the largest precision-and-recall gains are in Persons deprived
of their liberty (+271) and Persons with disabilities (+187).
The OHCHR Treaty Body Database lists 247 unique
E/C.12/<sess>/D/<n>/<yr>
symbols for CESCR Optional Protocol decisions
(sessions 55–79, 2014–2025). 147 are ingested with full
text via the docx-first pipeline (native DOCX from
documents.un.org; pre-2015 OLE
.doc converted via LibreOffice
so the <w:footnoteReference>
marker placement is preserved). The remaining
100 are catalog-only — every UN endpoint
returns 404, and OHCHR's own download surface answers
"Sorry there is no files available".
The pattern is structural rather than a pipeline gap:
88 of the 100 are bundled discontinuance
decisions where the Committee adopts a single
transmittal letter closing six to fourteen withdrawn
communications at once. Only the symbol appears in the
OHCHR catalog; no standalone document is published. The
remaining 12 are catalog rows with truncated titles
including one verifiable typo (E/C.12/77/D/82/018,
year missing the leading 2). Coverage is therefore
59 % full text, 41 % catalog-only.
Claude Sonnet 4.6 (Anthropic) was used as a coding assistant — proposing candidate patterns, reading samples of newly-applied labels to flag probable false positives, and writing the audit scripts. No paragraph label in the published corpus was generated by an LLM as free-form output. Every label is the result of a regular expression that a human can read, audit and modify; re-running the pipeline on the same input produces byte-identical output.
Suggested citation: Szoszkiewicz, Ł., & Kowalska, Z. (2026).
UNHRD — UN Human Rights Database. Zenodo.
https://doi.org/10.5281/zenodo.10495719
Citations of individual paragraphs should reference the original UN
document by its signature (e.g. CRC/C/GC/25 ¶12),
not this database.
Reference managers. Open a document in the reader and press the Zotero Connector's Save button: document pages carry citation metadata (title, issuing body, UN symbol as report number, adoption date, permalink), so the item lands as a Report rather than a web page. The Cite menu on every paragraph also copies CSL-JSON, RIS and BibTeX for Zotero, EndNote, Paperpile or Pandoc.
Check your citations. The Check tab takes a pasted passage and resolves every UN document symbol, “General Comment No. N” (with the committee named nearby), “Communication No. N/YYYY” and “X v. State” against this database, marks superseded texts and invented ones, verifies that a pinpointed paragraph exists and compares each quotation word by word with the source paragraph. It runs in the browser: the text is not sent anywhere; only the cited documents are fetched by identifier. It is an answer to citations invented by language models, which the Stanford RegLab studies measured in 17–33 % of answers from commercial legal tools.
The General Comments corpus is published as a paragraph-level
dataset on Hugging Face:
lszoszk/treaty-bodies-general-comments.
Every row is one paragraph from a Treaty Body General Comment or
General Recommendation, carrying its section path, footnotes (with
cross-reference resolutions), preamble flag, concerned-group labels,
and — new in the v3 release — cited_articles,
per-paragraph treaty-article references resolved to a specific
convention. 187 documents · 7,216 paragraphs · parquet, loadable
directly with the datasets library:
Citation graph. Every reference one document of the corpus makes to another — UN symbols, “general comment No. N”, “communication No. N/YYYY” in the text and footnotes — is extracted at build time into a document-level graph: graph.json (edges with the citing paragraph ids) and graph.csv (from, to, paragraphs). The reader shows the same data as “Cited by” and “Cites” in the right-hand drawer. It is text-derived: a case cited by name only, without symbol or communication number, is not linked.
New documents. Atom feeds list every addition, newest first: all, General Comments, decisions, Special Procedures, and one per treaty body or mandate (index). The changelog shows the same additions by day; returning visitors see a “new” chip on documents added since their last visit.
Related paragraphs. For every General Comment paragraph the reader and the dossier list the five most similar paragraphs of other General Comments (BGE-large embeddings, cosine similarity). Decisions and Special Procedures reports have no vectors yet, so the panel is General Comments only.
from datasets import load_dataset
ds = load_dataset(
"lszoszk/treaty-bodies-general-comments",
split="train")The dataset (curation, segmentation, annotation) is released under CC BY-NC-SA 4.0; the jurisprudence and Special Procedures collections are not yet part of the Hugging Face package — see the coverage notes for their status.
The corpus is also exposed as a
Model Context Protocol
server, so an LLM assistant (Claude, or any MCP-capable client) can
search and cite treaty-body paragraphs directly — returning the
verbatim text with its document signature and
paragraph number, instead of paraphrasing from training data. It
exposes search_paragraphs (full-text
search across the corpus, filterable by treaty body) and
lookup_by_citation (resolve a reference
like CRC/C/GC/25 ¶12 to its exact
paragraph), plus search_recommendations and
lookup_recommendation over the same UHRI records
the Recommendations scope shows. Source and setup:
github.com/lszoszk/mcp-unhrdb.
Software is released under the GNU Affero General Public License v3.0 — anyone running a modified copy as a hosted service must release their modifications under the same licence. The curated dataset (paragraph-level corpus, document metadata, section annotations, footnote cross-reference resolutions, concerned-group labels) is licensed under CC BY-NC-SA 4.0 — academic and non-commercial reuse is welcome with attribution; commercial use requires permission. The underlying UN documents remain under the United Nations' content terms. Built as a static site on GitHub Pages with vanilla JavaScript, FlexSearch and SheetJS. Set in Spectral, Source Serif 4 and JetBrains Mono.
Created by Łukasz Szoszkiewicz and Zuzanna Kowalska. Contact: l.szoszkiewicz@amu.edu.pl · source on GitHub. File issues, suggestions, missing documents and corrections through the GitHub repository.