Chinese Case Law API: Document Structure, Fields, and Integration Patterns for Legal AI Teams

Every legal AI team that expands jurisdiction coverage eventually hits the same wall: Chinese court judgments do not look like common-law opinions, and parsers tuned for US or UK reporters mis-segment them. China Judgements Online had published roughly 143 million effective judgments by December 2023, according to figures cited in SPC press briefings — the largest public collection of court decisions in the world.

This guide is a working anatomy of that corpus: the case-number (案号) system, the four-level court hierarchy, the cause-of-action (案由) taxonomy, and how the fixed sections of a PRC judgment map to the fields a Chinese case law API should expose.

It is written for the teams making integration decisions: legal AI product engineers, law-firm knowledge management, and data teams evaluating PRC court judgments as structured data.

Why PRC court judgments need their own data model

Two background facts shape every technical decision here. The first is scale: roughly 143 million judgments published through December 2023, with cumulative site visits above 108.1 billion, per SPC figures. The second is supply: annual publication fell from 19.2 million documents in 2020 to 14.9 million in 2021, 10.4 million in 2022, and about 5.11 million in 2023, per a December 2023 SPC press briefing; it partially recovered to more than 9.69 million in 2024 (up 92.7%) and about 10.98 million in 2025, per the March 2026 SPC work report — still only about 57% of the 2020 peak. Since January 2024, the complete national judgment database has been accessible only inside the court system.

Three engineering consequences follow:

• The corpus is finite, not streaming. You cannot assume a crawler will converge toward completeness. The archive you integrate is the archive you license on day one.
• Structure must be designed in. With publication patterns shifting year to year, consistent fields across the full archive are what keep retrieval quality stable.
• The same corpus serves retrieval and training. Chinese is about 1.3% of web content (down from 4.3% over eleven years) versus 59.8% for English, so a judgment corpus this size is among the densest high-quality Chinese legal text sources anywhere.

For the full supply curve and its market implications, see The state of Chinese case law data, 2026.

Anatomy of a Chinese court judgment: the segments that matter

PRC judgments are strikingly formulaic. Courts at every level follow essentially the same template — exactly what makes structuring 170M+ documents feasible, provided the data model understands the conventions.

Start with the case number (案号), printed in the header of every document. A schematic example, (2023)京01民终NNNN号, decodes as:

• (2023) — the year the case was docketed;
• 京01 — the court code: 京 for Beijing, 01 for the No. 1 Intermediate People's Court;
• 民终 — the case-type code: a civil matter (民) at second instance (终), i.e., an appeal;
• NNNN号 — the serial number for that docket year.

Case numbers are how Chinese lawyers and courts cite decisions; for a data team they are the native primary key, encoding court, region, year, case type, and procedural posture for free. Court codes sit on a four-level hierarchy — basic people's courts, intermediate courts, provincial high courts, and the Supreme People's Court, plus specialized IP, internet, maritime, and financial courts — so court level is a first-class ranking and filtering signal.

One more court-assigned label matters: the cause of action (案由), a hierarchical taxonomy maintained by the Supreme People's Court — a sales-contract dispute sits under the contract-disputes branch. Court-assigned rather than model-inferred, it is unusually reliable as a retrieval facet.

From anatomy to API fields: making 170M+ judgments machine-usable

A Chinese case law API earns its keep by exposing that anatomy as queryable fields rather than one undifferentiated text blob. The core schema:

Section-boundary segmentation is the highest-leverage field for retrieval quality. Fixed-window chunking mixes boilerplate party recitals into the same vectors as legal reasoning; chunking on 本院认为 boundaries yields embeddings that actually represent legal analysis.

Procedural linkage matters almost as much: an appeal judgment references its first-instance case number, so a structured corpus can collapse one dispute into a single timeline for deduplication and outcome-chain analysis.

Four integration patterns for a Chinese case law API

Across legal AI teams working with PRC court judgments as structured data, four patterns recur:

1. RAG retrieval layer. Hybrid retrieval beats pure dense search here: pre-filter by cause of action, court level, region, and date, then run embedding retrieval over reasoning sections only. Grounded answers cite case numbers users can verify — that loop is what builds trust with lawyer users.
2. Citator-style linking. PRC law has no stare decisis, so a US-style "cited by" graph is the wrong model. The useful edges are case-to-statute, case-to-case along procedural lineage, and alignment against the curated reference layers: SPC Guiding Cases and the People's Court Case Library, which opened to the public on February 27, 2024 with 3,711 cases and had grown to about 5,327 by January 2026. The curated layer is tiny next to a 170M+ corpus — the corpus supplies the statistical depth behind it.
3. Eval set construction. Structured metadata enables stratified sampling by cause of action, court level, region, and year for outcome-prediction, statute-retrieval, and fact-extraction benchmarks, with decision dates supporting leakage-free temporal splits. One caution: documents from the 2021–2023 publication trough are underrepresented on the public web, so web-scraped eval sets skew. A full snapshot lets you stratify against the actual distribution.
4. Agent tools and MCP. Agents need two primitives: deterministic fetch-by-case-number and parameterized search over the fields above. SinoVerdict exposes both through a REST API, bulk dataset, and MCP server, so MCP-compatible clients can call Chinese case law as a tool — see how access works.

Why raw HTML scraping fails — and what a structured snapshot solves

Teams sometimes budget China coverage as "a crawler plus a parser." Three realities argue against that plan:

• A crawl started today cannot reconstruct the corpus. Annual publication in 2025 ran at roughly 57% of the 2020 peak, and the complete national judgment database launched in January 2024 is restricted to internal court use. Documents are also removed over time. Whatever a new crawler collects will not be the 143-million-document archive.
• Public pages are display-oriented HTML with no machine-readable structure. Case-number parsing, court-code resolution, section segmentation, party-role tagging, and statute-citation extraction all have to be built, and conventions drift across years and thousands of courts. At this scale normalization is the real cost; fetching is the cheap part.
• There is no official free bulk channel to fall back on. When Harvey integrated eight jurisdictions' official free legal sources in July 2025 — Austria's RIS, Finland's Finlex, and India's Indian Kanoon among them — China was not on the list; no comparable PRC source exists. For jurisdictions without open infrastructure, per-jurisdiction licensing is the working model, as Legora's content agreements with Germany's Otto Schmidt and Denmark's Djøf Forlag illustrate.

A structured snapshot inverts the equation: documented coverage of 170M+ judgments through 2023, one normalized schema across the archive, provenance you can show procurement and counsel, and incremental updates available on request. For commercial structures and diligence questions, see the guide to licensing Chinese court data.

How PRC judgment data differs from English case law data

If your stack was designed around US or UK case law, five structural differences deserve attention:

Language is the sixth difference, and the decisive one. Primary PRC case law is in Chinese, and English-language coverage from incumbent providers is thin at corpus scale. As of mid-2026: LexisNexis's China law collection (LNCHNL) comprises roughly 3,000 English-translated documents; vLex — acquired by Clio for $1 billion in November 2025, claiming 110 jurisdictions and more than a billion documents — lists international collections and secondary commentary on its public China page, while mainland China does not appear in its primary-law jurisdiction list; and the Westlaw China site is currently unreachable.

The practical conclusion: at corpus scale you work in Chinese, joining on case numbers and statutes rather than precedent chains. Plan for cross-lingual embeddings or translate-at-query, statute-aware citators, and parsers that resolve 案号 formats.

Compliance notes for cross-border use (informational, not legal advice)

The following is informational background only, not legal advice. Engage qualified PRC counsel before structuring any data transaction.

• Published judgments are redacted public documents. Under Supreme People's Court publication rules, party names are retained while identifiers such as national ID numbers and home addresses are removed before release.
• PIPL distinguishes de-identification from anonymization. The distinction affects how party-related fields are treated downstream; conservative integrations apply field-level handling policies.
• Cross-border transfers follow defined pathways. PRC law provides three routes — security assessment, standard contractual clauses, or certification. A credible licensor should be able to explain which pathway its arrangement relies on and why.
• Data Security Law, Article 36. Providing data stored in China to foreign judicial or law-enforcement authorities requires prior approval, which is why license agreements typically exclude use for US discovery or law-enforcement production.

The buyer-side takeaway: usage scope belongs in the license text, compliance documentation in diligence — the licensing guide above lists the questions to ask.