Datasets:
Open India Law
Open, structured Indian primary law - plus the scrapers that build it. Every judgment of the Supreme Court of India and all 25 High Courts, the decisions of 15 tribunals and regulators, and Central, State and Union Territory legislation down to the individual section. Normalized to one schema, exclusively from official government sources.
| Volume | Period | |
|---|---|---|
| Court judgments | 12,848,644 | 1950 to 2025 |
| Tribunal and regulator matters | 813,168 | 1985 to 2026 |
| Enactments | 22,265 | 1806 to 2026 |
| Sections of legislation | 1,098,577 | individually searchable |
District and trial court decisions are not included.
Just want to read the documents?
Parquet is for pipelines. If you want to open an actual judgment or Act:
- Browse 22,264 Acts and regulations - pick a jurisdiction or a regulator, filter by title, click to open the PDF. Nothing to download.
- tribunals.vaquill.ai - tribunal and regulator decisions as PDFs, by forum, bench, party and date.
- news.vaquill.com - Indian courts, regulators and legal developments, readable rather than machine-readable.
- The raw file listing - every published file, if you want the Parquet directly.
- Every provision in the legislation and regulator files carries a
source_urlpointing at its own PDF, so you can go from a chunk straight to the document it came from. - Judgment source PDFs are on the AWS Open Data Registry (High Court, Supreme Court); we do not re-host them.
Courts
| Court | Judgments | Earliest | Latest |
|---|---|---|---|
| Supreme Court of India | 34,954 | 1950 | 2025 |
| Patna High Court | 1,615,041 | 1967 | 2025 |
| Bombay High Court | 1,595,948 | 1953 | 2025 |
| Allahabad High Court | 1,498,250 | 1992 | 2025 |
| Madras High Court | 1,494,952 | 1997 | 2025 |
| Telangana High Court | 1,004,138 | 1963 | 2025 |
| Kerala High Court | 916,190 | 1950 | 2024 |
| Karnataka High Court | 581,276 | 1998 | 2025 |
| Chhattisgarh High Court | 508,791 | 1970 | 2025 |
| Punjab and Haryana High Court | 483,253 | 2008 | 2025 |
| Gujarat High Court | 411,638 | 1982 | 2025 |
| Madhya Pradesh High Court | 402,632 | 2000 | 2024 |
| Rajasthan High Court | 324,567 | 1989 | 2025 |
| Delhi High Court | 322,940 | 1960 | 2025 |
| Gauhati High Court | 285,714 | 2000 | 2025 |
| Orissa High Court | 283,063 | 1992 | 2025 |
| Andhra Pradesh High Court | 254,482 | 1995 | 2025 |
| Jharkhand High Court | 246,374 | 1993 | 2025 |
| Calcutta High Court | 202,565 | 1960 | 2025 |
| Himachal Pradesh High Court | 184,175 | 1970 | 2025 |
| Uttarakhand High Court | 123,853 | 1950 | 2025 |
| Jammu and Kashmir High Court | 40,289 | 2003 | 2025 |
| Tripura High Court | 18,942 | 2013 | 2025 |
| Manipur High Court | 7,903 | 2017 | 2025 |
| Meghalaya High Court | 6,261 | 2010 | 2025 |
| Sikkim High Court | 453 | 2000 | 2025 |
Tribunals and regulators
Matters are distinct cases. Reasoned decisions is the estimated subset that is a substantive decision rather than a procedural order sheet, from a sample of 150 documents per forum - about 348,516 of 2,112,201 documents. Size this tier on that number, not the document count. Each forum links to the scraper that built it.
Tribunal material is indexed by case, not yet by full text. It is 2.1M PDFs with no text layer, of which roughly 348,516 are reasoned decisions, and extracting them is a large OCR and parsing job rather than a quick pass. Supporting scripts live in scripts/tribunals/, with per-forum schema in scripts/tribunals/schema/.
If you are building on this corpus and need that tribunal text, or the same extraction run over your own sources, we do that work under contract. We have put roughly 16 million Indian legal PDFs through this pipeline. Email contact@vaquill.ai.
Regulators
Each regulator ships as its own file, in_<body>_regulations.parquet, so you can take the
Securities and Exchange Board of India without pulling the Reserve Bank of India.
| Issuing body | Instruments | Provisions |
|---|---|---|
| Ministry of Corporate Affairs (MCA) | 2,666 | 46,480 |
| Reserve Bank of India (RBI) | 2,640 | 104,143 |
| Ministry of Environment, Forest and Climate Change (MOEFCC) | 1,399 | 116,227 |
| Directorate General of Foreign Trade (DGFT) | 1,233 | 4,252 |
| Securities and Exchange Board of India (SEBI) | 1,144 | 88,310 |
| Telecom Regulatory Authority of India (TRAI) | 834 | 22,720 |
| Central Pollution Control Board (CPCB) | 558 | 13,752 |
| Ministry of Law and Justice (LAW) | 558 | 77,209 |
| State GST and tax authorities (TRIB) | 354 | 41,708 |
| Insurance Regulatory and Development Authority of India (IRDAI) | 314 | 16,238 |
| Department of Financial Services (DFS) | 283 | 18,509 |
| Central Board of Indirect Taxes and Customs (CBIC) | 169 | 4,806 |
Legislation
Sourced from India Code. Every section is held and indexed separately.
Scrapers: Acts crawler · HTML section extractor · repealed Acts · gap re-scrape · amendment metadata · normalization pipeline
| Jurisdiction | Enactments | Sections | Earliest | Latest |
|---|---|---|---|---|
| Central | 13,720 | 628,863 | 1834 | 2026 |
| Assam | 987 | 12,429 | 1806 | 2020 |
| West Bengal | 715 | 32,238 | 1947 | 2024 |
| Rajasthan | 383 | 26,759 | 1860 | 2025 |
| Maharashtra | 379 | 30,271 | 1866 | 2025 |
| Chhattisgarh | 359 | 25,368 | 1860 | 2023 |
| Uttar Pradesh | 338 | 14,308 | 1885 | 2025 |
| Kerala | 322 | 19,164 | 1951 | 2025 |
| Tamil Nadu | 308 | 15,524 | 1864 | 2023 |
| Punjab | 303 | 18,037 | 1887 | 2025 |
| Bihar | 298 | 10,910 | 1894 | 2024 |
| Odisha | 282 | 11,370 | 1908 | 2025 |
| Karnataka | 279 | 29,485 | 1899 | 2025 |
| Telangana | 259 | 16,785 | 1837 | 2024 |
| Chandigarh | 256 | 21,805 | 1860 | 2018 |
| Uttarakhand | 244 | 10,531 | 1901 | 2022 |
| Gujarat | 226 | 12,868 | 1867 | 2025 |
| Madhya Pradesh | 216 | 17,477 | 1860 | 2023 |
| Haryana | 211 | 11,787 | 1897 | 2025 |
| Nagaland | 202 | 4,861 | 1954 | 2024 |
| Himachal Pradesh | 194 | 10,765 | 1952 | 2023 |
| Andhra Pradesh | 190 | 12,264 | 1954 | 2023 |
| Manipur | 187 | 4,560 | 1924 | 2023 |
| Dadra and Nagar Haveli | 169 | 15,619 | 1860 | 2018 |
| Goa | 154 | 14,539 | 1867 | 2024 |
| Jharkhand | 137 | 10,831 | 1887 | 2021 |
| Jammu and Kashmir | 137 | 9,223 | 1945 | 2020 |
| Tripura | 121 | 8,070 | 1926 | 2023 |
| Arunachal Pradesh | 115 | 6,996 | 1891 | 2025 |
| Sikkim | 111 | 4,266 | 1975 | 2022 |
| Meghalaya | 105 | 5,550 | 1970 | 2024 |
| Delhi | 96 | 8,006 | 1870 | 2019 |
| Mizoram | 83 | 2,015 | 1988 | 2023 |
| Puducherry | 78 | 7,117 | 1897 | 2019 |
| Ladakh | 55 | 3,119 | 1945 | 2018 |
| Andaman and Nicobar Islands | 45 | 4,626 | 1894 | 2016 |
| Lakshadweep | 1 | 171 | 1995 | 1995 |
Parliament
Lok Sabha debates, Law Commission reports and gazette records: download-lok-sabha-debates.py.
Document pipeline
Indian courts and tribunals publish PDFs, very often scanned with no text layer. Turning those into clean, section-aware, citable text is the hard part, and it is what this project actually contributes.
| Stage | Script |
|---|---|
| PDF to markdown (text layer) | pymupdf4llm_parser.py |
| Unified parser across document kinds | unified_legal_parser.py |
| Section boundary detection | legal_section_detector.py, v2 |
| Section-aware chunking | legal_chunker.py |
| Metadata extraction | metadata_extractor_v3.py |
| Parallel chunking driver | run-chunking-parallel.py, full pipeline |
| OCR fallback (no text layer) | ocr-mistral.py |
| Chunk schema | UNIFIED_CHUNKING_SCHEMA.md |
| Read AWS Open Data HC parquet | aws-hc-parquet-reader.py |
A court-structure-aware variant is in scripts/pipeline/court/; chunking tests in scripts/pipeline/tests/.
What you get (output format)
Scrapers write JSONL - one normalized record per line - to $OUT_DIR (default ./data), plus source PDFs where the forum publishes them.
No database, no cloud storage, no credentials required.
Legislation records:
| Field | Meaning |
|---|---|
act_id |
Stable identifier, e.g. IND_central_1860_45 |
title |
Short title of the Act |
chapter / section_number |
Position in the Act's hierarchy |
section_title / text |
Heading and text of the provision |
act_status / section_status |
in_force, repealed, spent |
state |
Jurisdiction (central or the State) |
year / amendment_count |
Enactment year, number of recorded amendments |
source_url |
Back-link to the authoritative government page |
Tribunal records:
| Field | Meaning |
|---|---|
case_id |
Stable dedup key |
case_number / title |
Registry number and cause title |
bench / judges |
Bench and coram |
decision_date / year |
Date of the order |
doc_type / is_judgment |
Order, judgment, notice |
source_pdf_url |
Back-link to the tribunal's own PDF |
Download
from datasets import load_dataset
judgments = load_dataset("vaquill/open-india-law", "judgments", split="train")
acts = load_dataset("vaquill/open-india-law", "legislation", split="train")
regs = load_dataset("vaquill/open-india-law", "regulations", split="train")
kerala = load_dataset("vaquill/open-india-law", data_files="in_kerala_judgments.parquet")
sebi = load_dataset("vaquill/open-india-law", data_files="in_sebi_regulations.parquet")
huggingface.co/datasets/vaquill/open-india-law · mirror at oss-data-in.vaquill.ai
Snapshot v2026.08: 26 judgment files (32,572,660 chunks, 53.6 GB), 37 legislation files
and 12 regulator files (1,098,269 provisions between them). Judgment rows are one per chunk - group by case_id and order by
chunk_index to reassemble a judgment.
You do not need to run any scraper below to use the data. They are here so the corpus is reproducible and auditable.
Vector embeddings
Every chunk is published with its embedding, as Qdrant per-shard snapshots.
| Collection | Points | Shards | Size | Content |
|---|---|---|---|---|
legal_corpus_v1 |
19,595,718 | 4 | 272.8 GB | High Court and Supreme Court judgment chunks |
legal_corpus_v2 |
11,823,753 | 4 | 179.6 GB | Tribunal and regulator decision chunks |
acts_india |
1,098,577 | 2 | 11.2 GB | Legislation and regulatory provisions |
32,518,048 vectors, 463.6 GB, taken from Qdrant 1.16.3. Embeddings are Voyage AI voyage-4 series, 1024 dimensions, cosine distance. Each collection also carries a named sparse vector used for BM25 hybrid search.
Embed your queries with the voyage-4 series. Vectors from a different model live in a different space, so similarity scores against them are not meaningful.
Import into Qdrant
Snapshots are per-shard, so create the collection first with a matching shard_number, then recover each shard.
Qdrant fetches each snapshot itself and verifies the published SHA256 before accepting it.
# 1. create the collection
curl -X PUT http://localhost:6333/collections/legal_corpus_v1 \
-H 'Content-Type: application/json' -d '{
"shard_number": 4,
"vectors": {"dense": {"size": 1024, "distance": "Cosine", "on_disk": true,
"quantization_config": {"scalar": {"type": "int8", "quantile": 0.99}}}},
"sparse_vectors": {"sparse": {}}
}'
# 2. recover each shard straight from the mirror
BASE=https://oss-data-in.vaquill.ai/qdrant/legal_corpus_v1
for N in 0 1 2 3; do
SNAP=$(curl -s "$BASE/shard-$N/index.json" | jq -r .snapshot)
SUM=$(curl -s "$BASE/shard-$N/$SNAP.checksum")
curl -X PUT "http://localhost:6333/collections/legal_corpus_v1/shards/$N/snapshots/recover" \
-H 'Content-Type: application/json' \
-d "{\"location\": \"$BASE/shard-$N/$SNAP\", \"checksum\": \"$SUM\", \"priority\": \"snapshot\"}"
done
Use shard_number: 2 for acts_india.
Manifest of every shard, size and checksum: qdrant/index.json.
Full guide including verification and disk requirements: QDRANT_RESTORE.md.
Quick start
pip install -r requirements.txt
npm install # the TypeScript scrapers
cp .env.example .env # optional - only proxies and OCR need keys
# Legislation: India Code (Central + State Acts)
OUT_DIR=./data npx tsx scripts/legislation/indiacode-scraper.ts
# A tribunal (the Income Tax Appellate Tribunal):
OUT_DIR=./data npx tsx scripts/tribunals/itat-scraper.ts
# Turn scraped PDFs/HTML into normalized section records
OUT_DIR=./data python scripts/legislation/pipeline/run-pipeline.py --help
Every script is self-documenting - run it with --help, or read its module docstring for the exact source and options.
Sourcing and provenance
Government-only. Every record traces to a court, tribunal or government publisher, and keeps the source URL it was ingested from. No commercial law reporter and no third-party aggregator material enters this corpus.
Not every official Indian body publishes on a .gov.in domain, so the source list is not
filterable by suffix alone. These are the official portals of the bodies named: the Reserve
Bank of India on rbi.org.in, the Pension Fund Regulatory and Development Authority on
pfrda.org.in, the Delhi Real Estate Regulatory Authority on erera.co.in, and the Bar
Council of India's All India Bar Examination on allindiabarexamination.com.
Scrapers identify as VaquillLegalBot/1.0 and respect each site's robots.txt.
The source judgment PDFs are not republished here. They are already public under CC BY 4.0 on the AWS Open Data Registry - High Court (~17.8M judgments, ~1.25 TiB) and Supreme Court, published by Dattam Labs. Re-hosting 1.36 TB of already-public files would add nothing, so this project publishes the layer they lack: extracted text, chunking, provenance and metadata.
The embeddings ARE published, as Qdrant snapshots. See Vector embeddings. Still out of scope: the citation graph, which is coupled to our own infrastructure.
What has been removed, and why
Identifying detail that a statute bars from publication. Indian courts anonymize sexual-offence victims in the large majority of cases, so this dataset does not exclude by statute category: doing so would remove roughly 180,000 judgments that are lawful to publish precisely because the court controlled the disclosure. What is redacted is the narrow band where identity actually leaks:
- names of relatives of victims and of protected children, where the court named them
- telephone numbers and email addresses
- Permanent Account Number, Indian Financial System Code and Aadhaar numbers, each only where the surrounding text identifies the number as such
That last condition matters. Those patterns are shape-only and Indian case numbers share the
shapes: WPCT0123456 is an Indian Financial System Code by shape and ABCDE1234F is a
Permanent Account Number by shape. Masking on shape alone corrupts the judgment text, so a
context word is required nearby.
Documents a court directed not be published, including in-camera matters. Those sit outside the statutory exception this corpus relies on.
Counts of what each rule removed are published with each snapshot.
Reporting a problem
If this corpus contains material that should not be public, write to contact@vaquill.ai
with the subject line Open India Law - redaction request. Directions of Indian courts and
tribunals are honoured. Because a published snapshot is a fixed artifact with published
checksums, corrections are made by publishing a superseding snapshot rather than editing one
in place.
Important caveats
1. Several sources serve Indian traffic only. India Code in particular returns "The specified URL is inaccessible at this time" to non-Indian IPs while serving its homepage normally.
Run those scrapers from an Indian host, or configure a proxy in .env.
If a run returns almost nothing and you are outside India, that is the usual cause.
2. Some scripts will stop working over time. These target live government websites, which get redesigned, move URLs, change HTML, or add anti-bot measures. A scraper that worked at publish time can break later. That usually needs a small parser update, not a rewrite. Please open an issue or PR - fixes to individual parsers are exactly where community help compounds.
3. A browser is needed for some forums. Most sources are plain HTTP.
A handful render via JavaScript and use Playwright - you will need its browsers installed (npx playwright install).
4. The tribunal corpus is PDF-only. These forums publish scanned or generated PDFs with no machine-readable text. Text extraction is a separate step and its quality varies by forum.
5. Snapshots are point-in-time, not current law. Indian legislation changes continuously and tribunal orders are appealed. Output is an archive as of the run date - always verify against the official source before relying on it. This is not legal advice.
Licensing and commercial use
- Scripts - Apache-2.0 (
LICENSE). Free, including commercial use. - Data / compilation - CC BY 4.0 (
data/LICENSE.md). Free with attribution. - The underlying legal text - a Government work under s.17(d) of the Copyright Act 1957, reproducible under s.52(1)(q). We hold no rights in it and grant none; your right to use it comes from the statute.
The dataset is free for everyone. You never need to email us or ask permission to use it, including for commercial products, as long as you attribute it.
Attribution is a condition our sources impose, not a preference of ours - the eCourts policy and those of the National Company Law Tribunal and the National Company Law Appellate Tribunal all require the source to be prominently acknowledged. That is why this corpus is CC BY rather than CC0.
Want it built for you?
The dataset is free and always will be. Separately, we build legal data pipelines under contract: bulk extraction from scanned or non-machine-readable sources, OCR at scale, normalization into a schema you can query, and ongoing refresh.
That is what produced this corpus. If a law firm, publisher or legal-research product needs data at this scale and cannot get there from the raw sources, email contact@vaquill.ai.
Contributing
New-source parsers, coverage fixes, and - especially - repairs to scrapers that broke when a government site changed are welcome. Open a PR against the relevant script in the tables above.
Particularly useful right now: verified copyright/terms pages for the forums marked unverified in coverage.yml, captured from an Indian IP.
Most of those pages are unreachable from outside India, and each one unblocks a corpus.
Particularly useful: verified copyright and terms pages for the forums still marked
unknown in coverage.yml, captured from an Indian IP. Most are unreachable
from outside India, and each one unblocks a corpus.
Maintained by
Vaquill AI. Full measured coverage, including year-by-year tables and the held-vs-embedded reconciliation, is in COVERAGE.md.
Questions, ideas, or want to help? DM me on LinkedIn.
The law is public. Making it usable should be too.
- Downloads last month
- 158