How this case-law search works — and how well
A plain-English note for lawyers. Last reviewed October 2026.
What it searches
Every English-language ECtHR judgment of the corpus, with each day's new judgments added by the daily update — about 20,000 judgments and Grand Chamber decisions, ~1.33 million numbered paragraphs — restricted to substantive text (reasoning, facts, admissibility, just satisfaction, separate opinions). Pure boilerplate (headers, annexes, one-line summaries) is excluded.
How it works (without the jargon)
- Meaning, not keywords. Each paragraph becomes a numeric "meaning fingerprint" (an embedding); your query is fingerprinted the same way and matched by legal substance. That is why it understands plain language — "can the police read my texts?" finds the Article 8 interception case-law even though those words never appear.
- Authority weighting. Results are nudged toward the Court's leading cases (HUDOC "Key case"/importance), because a practitioner wants the authoritative judgment.
- A second read. The top 100 candidates are re-read against your query by a separate model (Voyage rerank-3) to sharpen the order.
How we measured it
We did not grade ourselves. We used the Court's own 41 official Case-Law Guides as an answer key — each states a legal proposition and pinpoints the exact case paragraphs that support it (~8,400 propositions, ~24,000 pinpoint citations) — and checked whether the system returns the cited case and the cited paragraph.
A second, larger test — 1,570 cases that carry an official summary — checks retrieval at scale, including a plain-language track that a layperson (not a lawyer) would type.
How well it works
| On the Court-Guides benchmark (409 items) | First | Top 10 |
|---|---|---|
| Correct case retrieved | ~64% | ~89% |
| Exact cited paragraph retrieved | ~39% | ~75% |
Judgments only in French (the option). On 400 propositions of the French editions of the Guides whose cited judgment exists only in French, ticking Also judgments only in French puts that judgment first for ~37% and in the top 10 for ~79% (the exact paragraph in the top 10: ~61%); without it they cannot be found. It costs the English results a little (cited case in the top 10 on the Guides benchmark: 88.0% → 86.6%), which is why it is off by default.
An independent AI-lawyer review found ~85–90% of the top-10 paragraphs genuinely relevant — higher than the "exact paragraph" figure, because the Guides cite only some of the relevant paragraphs, so the strict measure understates real usefulness.
Where to be cautious
- A research aid, not legal advice, and never a substitute for reading the judgment. Always open the case on HUDOC and verify.
- Relevant ≠ current law. A retrieved case may have been refined, distinguished or overtaken by later case-law. Check how the principle stands today.
- "The paragraph" is rarely one paragraph. The Court states a principle, applies it, and concludes across several §§; the system may surface the reasoning where you expected the holding. Read around the hit (click a paragraph to expand neighbours).
- It favours leading cases — a niche but on-point lesser-known judgment may rank lower.
- Absence is not proof. If a point is not in the results, that is not evidence it does not exist — refine the query.
- Coverage. The English judgments of the corpus (new ones are added every day), and, when you tick Also judgments only in French, the authentic French texts of the judgments HUDOC publishes only in French; excludes decisions (other than 31 Grand Chamber ones), non-judgment documents and the machine translations.
Examples
Semantic Search
Experimental. Semantic Search runs on its own index, which the daily update extends with each new judgment: about 20,000 English-language judgments and Grand Chamber decisions (about 1.33 million paragraphs). It does not cover the machine translations of judgments that exist only in French, or decisions other than the Grand Chamber ones. Tick Also judgments only in French to add the authentic French texts of the judgments HUDOC publishes only in French; an English query finds them too. For the complete, current corpus use Search. How Semantic Search works, and how well →
Any language. The system rewrites your text into a dense legal query and detects the relevant article(s).
Advanced options
Six exemplary cases
The court understood you as saying
The cases that answered your query
X axis = year of judgment · Y axis = relevance, relative to the other results · node size = times cited (log scale). Arrows show citations between cases visible in this view — a selected case may also cite or be cited by cases not shown here; check the dossier to see all of them. Foundational doctrines (Klass-style) become large; recent cases stay small.
⚐ Chart capped at the 30 most relevant judgments. The full list is in “Ranked paragraphs” on the left.
Your results over time
The judgments that answered your query, placed by year of judgment.
Each dot is one case, placed by year of judgment. Bigger dot = more matching passages · labelled dots are the closest matches to your query · highlighted = what you're reading. Click a dot to read that judgment; click elsewhere in a year's column, or a year in the strip below, to read every matching paragraph from that year.