What is this search engine?
This is an AI-powered legal search engine for European and Hungarian court case-law. It does not search for keywords like a traditional database: it understands the meaning of your question, finds the most relevant decisions, and answers in natural language, backed by verbatim quotations. By default it searches the Court of Justice of the European Union (CJEU) and the European Court of Human Rights (ECtHR); you can also add the Hungarian Constitutional Court, the Hungarian ordinary courts, the Hungarian Public Procurement Arbitration Board, the EFTA Court and the Kúria's non-binding guidance from the selector under the search box — seven sources in all.
Judgments and orders of the Court of Justice and the General Court, and Advocate General's Opinions, updated automatically every night.
Decisions of the Hungarian Constitutional Court from 1990 to the present, updated daily. Hungarian-language source material.
The public collection of Hungarian ordinary-court decisions (district court to the Curia), supplemented with Curia decisions missing from it that the Curia publishes on its own website (election, referendum, local-government and other special proceedings). Hungarian-language source material.
Judgments and advisory opinions of the EFTA Court, updated weekly — an optional source.
Decisions and orders of the Hungarian Public Procurement Arbitration Board (KDB), continuously from 2018 (a few dozen earlier), updated daily — an optional source. Their judicial review is linked into the decision chains. Hungarian-language source material.
College opinions and positions, case-law analyses, old guidelines and principled decisions of the Hungarian Curia (kuria-birosag.hu), updated daily. Professional guidance — not court decisions and not binding; the binding uniformity decisions are among the ordinary-court decisions, marked. Hungarian-language source material.
From the HUDOC database: Grand Chamber, (pre-1998) Plenary Court, Chamber and Committee judgments, decisions (e.g. inadmissibility, strike-out) and advisory opinions. The original text is English or, where there is no English version, French (quotes stay in the original language; a machine translation is offered). For some cases HUDOC also publishes an unofficial Hungarian translation; the search does not currently display these. Last update: —.
A decision (e.g. inadmissibility, strike-out) has its type marked separately. Separate opinions and the parties' submissions are also marked and are not the Court's position.
How does it search? (step by step)
- Indexing. Each decision is split into paragraphs and labelled by section (header, operative part, background, the court's reasoning, closing part). This lets the engine know whether a sentence comes from the facts or from the substantive legal reasoning.
- Semantic embedding. Every passage is turned into a 1024-dimensional vector with Voyage AI's model. Texts with similar meaning end up close together — so the search is on meaning, not on the exact word.
- Hybrid retrieval. Your question is turned into a vector too, and we search in parallel by meaning (vector similarity), by literal match (lexical) and by case number (if you give one, that decision is pinned in). The results are merged.
- Reranking. A separate, finer model (Voyage rerank-3-lite) reviews the best candidates from every selected court together and brings the ones that fit your question most precisely to the front.
- Pulling in the decision chain. For every relevant hit, the engine attaches the full appeal chain the decision belongs to (see the next section).
- Answering. Selected passages of the most relevant decisions — together with a summary of their chains — are handed to OpenAI's language model, which, based on these and only these, writes the answer: a summary, verbatim quotations and a short analysis, noting where the courts agree or diverge.
Important: the AI does not read the full decisions
The answering model works from at most 10 decisions (at most 6 from any one court, and at least 2 from every selected court if it has enough relevant hits), and from these it receives at most 3 passages per decision (a passage is roughly one or two paragraphs) — not the full text. The passages are chosen by the reranker as the ones that best match your question.
So it can happen that for a decision the model sees only the facts, not the reasoning or the operative part — if the facts were what matched your question most closely. In that case the answer cannot say anything certain about that decision's outcome or reasoning.
You can always open the full text yourself: click the decision in the result list or at the quotation, split by section. On an important question, always read the operative part and the reasoning too.
Decision chains: the full journey of a case
A legal dispute rarely ends with a single judgment: the first-instance decision is appealed, the appellate judgment is reviewed by the Curia (the Hungarian supreme court), a constitutional complaint may follow at the Constitutional Court, and during the proceedings the court may have referred questions to the CJEU for a preliminary ruling. In the public collections these are separate documents — nothing shows they belong together. This engine automatically links them into chains: it currently tracks more than 100,000 chains, rebuilt daily. Decisions of the Public Procurement Arbitration Board are linked in together with their judicial review.
How do we know what connects to what? From two sources that cross-check each other. First, from the operative part of the higher court's decision: there the court names the case number of the challenged decision verbatim (“the Court upholds judgment No. 3.Kf.650.056/2019/5 of the Budapest-Capital Regional Court”) — we extract this from the text itself. Second, from the official metadata of the court repository. Where the two agree, the link is highly reliable; where only one source asserts it, the system treats it as a weaker link. Constitutional Court decisions name the challenged court decisions themselves; the CJEU link is based on the Court's official register of preliminary references.
Chains are shown in chronological order — oldest decision first, newest last. Note that a CJEU preliminary ruling is not an appeal: it is an answer given during the Hungarian proceedings to a question referred by the Hungarian court, so it appears between the Hungarian decisions in the timeline, not “on top” of the chain. Chain members drawn with a dashed border are decisions we know exist (the metadata refers to them) but whose text is not in the collection.
Every decision is framed with a colour according to whether it survived review by the higher forum:
- green — it survived: the higher forum upheld it, or the Constitutional Court rejected/dismissed the complaint against it (so the challenged decision lived on);
- red — it was overturned: changed, quashed or annulled;
- yellow — uncertain: partially modified, or the outcome cannot be established unambiguously;
- no colour — the top of the chain (nothing reviewed it) and the CJEU (nothing can review it).
Safety rule: if anything higher up in the chain was overturned, the decisions below it are marked red even if they were not directly affected — because in that case it cannot be known with certainty what remained in force. The exception is a lower decision that a surviving higher decision explicitly upheld. In doubtful cases the system errs on the side of caution, not confidence.
Why does this matter? The search may find the relevant passage at any point of the chain — the reasoning of a lower court can be valuable even if it was later overturned. The colour tells you at a glance whether what you are reading is still good law. The answering model receives the chain too: case numbers in the written answer carry the same coloured frames, and the model flags it when a cited decision was later overturned.
How do you know the quotations are authentic?
Language models sometimes “make up” text. Here a machine check rules this out — it does not leave authenticity to the model:
- Every quotation in the answer is checked word for word against the original text of the decision (the passages the model was given). The check only tolerates differences in upper/lower case, typographic marks (e.g. quotation marks, dashes) and whitespace — the words and their order must match exactly.
- If a quotation cannot be found exactly in the source, it is dropped — it never appears on the page.
- Each quotation that is shown is linked to its source decision and the section it comes from, so it can be traced back.
This check is code-based and model-independent: the system itself computes the text match; the AI does not judge it. The check always runs on the original-language text; only after that, a quotation whose language differs from the interface (e.g. a Hungarian or French quote here) is shown in machine translation, with a “machine translation” badge and the original one click away.
What else the interface can do
- Choose which courts to search. Use the selector to include or exclude the CJEU, the Hungarian Constitutional Court, the Hungarian ordinary courts, the Public Procurement Arbitration Board the EFTA Court and the European Court of Human Rights. By default the CJEU and the ECtHR are selected on the English interface.
- Full decision text, split by section. Click any decision to open its full text, where every paragraph is labelled (header, operative part, background / facts, the court's reasoning, closing part). Click the labels at the top to jump to that section.
- Judge filter. You can narrow results by judge-rapporteur or panel member.
- Law filter (Hungarian ordinary courts). Narrow to decisions citing a given act — the old and the new code (e.g. the old and new Civil Code) are separate items.
- Quotes ↔ analysis balance. A slider sets whether the answer contains more quotations or more analysis.
- Weekly question allowance. The number of questions per week is limited; the interface tells you when you are close to the limit, and you can ask for more.
- Your past questions. Your earlier questions are kept on the left; clicking one reloads the saved answer (without regenerating). You only see your own questions.
Reliability — honestly
You should know how much to trust each element. The system is built largely automatically, and we aim to flag its errors rather than hide them:
- Quotations are machine-verified word for word — this is the system's strongest guarantee. The summary and the analysis, however, are the work of a language model: they can be wrong, and the model only sees selected passages of the decisions (see above), so the quotation and the original decision are always what counts.
- Decision chains are built by automated text analysis plus repository metadata. Roughly half of the links are confirmed by both sources independently (these are highly reliable); links that come from metadata only are weaker and are marked as uncertain in the interface. We also run a chronology check (an appeal cannot predate the decision it challenges): about 1% of links are provably wrong and are flagged as uncertain — and undetected errors may also exist.
- The green/red/yellow status is computed with deterministic rules from the operative part of the higher decision. Where the outcome is not unambiguous we deliberately mark it yellow — about one in six chain members is. Doubtful cases are pushed towards red (the safety rule), so red is sometimes stricter than reality; green marks accordingly carry higher confidence.
- Dates are day-precise for about 87% of decisions (read from the closing part of the decision text); for the rest only the year is known — there the ordering is decided by the appeal logic, not the date.
- The collection is not complete: not every Hungarian court decision is published (publication is selective and anonymised), so some chains have gaps — missing members are shown with a dashed border and we make no claim about their fate. Very recent, unpublished decisions are not in the system.
- Section labelling (operative part, reasoning, etc.) is machine-learned and can err at the level of individual paragraphs.
- Long chains (five or more members) are additionally reviewed by a separate AI check to confirm they really belong to one case — but with complex, remanded proceedings a wrong merge or split can still occur.
Machine access: use it from your AI assistant (MCP)
The search engine is not only available through this interface: your own AI assistant (e.g. ChatGPT, Claude or Codex) can query it directly over the standard MCP connection at https://dontes.app/mcp. You sign in with your own dontes.app account (OAuth); no separate permission is needed — the connection steps are on the Machine access page.
- Search — 10 relevant decisions by default, up to 50, with passages, the decision's metadata (court, case number, date, panel, judge-rapporteur, cited legislation, etc.) and its decision chain.
- Full text — any decision (a case number also works), with a data sheet, the chain members and where the operative part, the reasoning and the facts start in the text.
- Ready answer — dontes.app's own answer with verified quotations (from at most 10 decisions, as on the interface). Only a few per day: the assistant can do the analysis itself from the search results and the full texts.
Calls count against a daily allowance per user. Note: with machine access the full decision text goes to your assistant; writing the answer — and its accuracy — is then the assistant's responsibility. The verbatim quotation check applies only to dontes.app's own answer.
What it is good for — and what it is not
Works well for
- Mapping the case-law on a question
- Quickly finding relevant decisions
- Following a case across instances (did the decision survive review?)
- Comparing the CJEU and the Hungarian courts on a topic
- Collecting verbatim citations
- Pulling up a specific decision by case number
- Orientation, research, idea generation
Use with care / not suitable for
- It does not replace legal advice or a lawyer's judgment
- The summary and analysis are indicative — what matters is always the quotation and the original decision
- Chain links and the colour marks are automated — verify in the original decision before citing
- The collection may not be exhaustive for every type of case
- It does not know very recent, unpublished decisions
- In a concrete matter, always verify the original source
Curia guidance — what it contains and how much it counts
The Hungarian Curia's non-binding professional guidance, as a separate source: opinions and positions of its colleges, summaries of its case-law analysis groups, and old guidelines and principled decisions (source: kuria-birosag.hu, updated daily). These are not court decisions. The binding uniformity decisions are not here but among the ordinary-court decisions, marked ⚖. What is worth citing is best shown by what the courts themselves cite in their reasoning — so the table below gives that too. Hungarian-language material.
Contents
| Type (count) | Binding? | Citing decisions* | of which Curia · since 2018 |
|---|---|---|---|
| College opinion (171) | No — per the Curia, not binding "even on the courts" (Courts Act s. 27) | 16,051 | 6,231 · 10,202 since 2018 |
| College position (144, old) | No (never was); uniformity decisions 1/2014 PJE etc. list which still apply | 9,642 | 1,629 · 4,424 since 2018 |
| Case-law analysis (76) | No — summary opinion (Courts Act ss. 29–30) | 530 | 133 · 456 since 2018 |
| Guideline (1) | Formerly yes (Act IV of 1972 s. 48); now "applicable" (Courts Act s. 195(2)) | 1,536 | 218 · 136 since 2018 |
| Principled decision (6) | Formerly yes; now "applicable" (Courts Act s. 195(2)) | 961 | 209 · 206 since 2018 |
| Presiding judges' position (8) | No | 103 | 25 · 40 since 2018 |
For comparison (among the ordinary-court decisions):
| Type (count) | Binding? | Citing decisions* | of which Curia · since 2018 |
|---|---|---|---|
| Uniformity decision | YES (Courts Act s. 42(1)) | 8,055 | 2,465 · 5,162 since 2018 |
| Published decision (BH, EBH, EBD) | Precedent: the Curia's published decisions | 51,533 | 16,871 · 26,461 since 2018 |
What we deliberately left out
| Type (count) | Binding? | Citing decisions* | of which Curia |
|---|---|---|---|
| Positions of the national conference of heads of civil colleges (CKOT, 593) | No — "not binding on the courts" | 75–164 | 5–19 |
| Civil Code advisory panel opinions (35) | No — and not the Curia's position | 95 | 4 |
| Civil Procedure consultative panel (82) | No | 48–70 | 5–9 |
| Consultative panel minutes (11) | No | 15 | 3 |
Courts rarely cite these, and when they do they often treat them as weak. The Curia held that they "may assist adjudication … but have no binding force" (Pfv.20666/2021/6); a 2024 judgment rejected reliance on a CKOT position as an unauthorised statement (Courts Act s. 27/A; Mf.31032/2024/6); another held a consultative-panel minute "not binding on the court" (Gf.40426/2020/3). As the aim is to find sources that persuade a court, we left them out.
Good to know
- The college opinion is the most persuasive non-binding source: more decisions cite it in their own reasoning (16,051) than cite numbered uniformity decisions (8,055).
- Any note the Curia attaches to an item (e.g. "applicable only in cases under the old Civil Code") is always shown with the hit and passed to the answering model.
- 227 college documents also appeared in the ordinary-court database (BHGY) as if they were decisions, often in an old version (as of 8 November 2026). They now appear only here, with the Curia's current text.
- 91.8% of the courts' citations of numbered opinions and positions point to an item that is included here (27,998 of 30,506). The most cited missing items: MK 95, KK 31, KK 34, PK 10, BK 155. They are not available on the Curia's website, so they are not here either.
* Number of decisions citing the type in the court's own reasoning (not in the parties' arguments or the procedural history), out of the 246,163 published Hungarian ordinary-court decisions (authority documents excluded). The figures come from pattern matching; "reasoning" is our automatic section labelling; generic phrases may also catch courts of appeal's own college opinions; where two figures are given, the first counts numbered citations only, the second includes generic mentions. The item counts of the omitted sources are as of 8 November 2026. Measured: 20 November 2026; updated weekly.
Technical details
For power users and developers — how the system is built, its tools and parameters.
| CJEU collection | EUR-Lex / Formex (judgments and orders of the Court of Justice and the General Court, Advocate General's Opinions); German language version where available |
| Hungarian collections | Constitutional Court (mkab.hu / alkotmanybirosag.hu) and the public ordinary-court repository (eakta.birosag.hu) |
| Curia supplement | kuria-birosag.hu decision lists (elections, referendums, local-government norm control, cases following Constitutional Court decisions, freedom of assembly, preliminary references). Only decisions not in the ordinary-court repository are added (~800, mostly 2006–2019 election cases). A duplicate means the same case number and at least 90% identical characters; if the repository publishes it later, its copy is kept. Press summaries are not included. |
| EFTA Court | eftacourt.int (judgments and advisory opinions) |
| ECtHR collection | hudoc.echr.coe.int (Grand Chamber, Chamber and Committee judgments, admissibility decisions, advisory opinions; automatic updates) |
| Public Procurement Arbitration Board | dontobizottsag.kozbeszerzes.hu (decisions and orders, continuously from 2018) |
| Updates | systemd timers (Europe/Budapest): CJEU 02:00, ordinary courts 03:00, Constitutional Court 04:00, Curia supplement 04:30, Procurement Board 05:00 daily; EFTA Court Sundays 03:00; decision chains rebuilt every morning. Delta logic by publication / indexing date; completeness of the ordinary-court mirror measured by two independent weekly checks |
| Vector database | Qdrant 1.12 (HNSW index, on-disk payload) |
| Embedding models | Voyage voyage-4-large (CJEU, ordinary courts, EFTA Court, Procurement Board) and voyage-3-large (Constitutional Court), 1024 dimensions |
| Reranker | Voyage rerank-3-lite (cross-court, source-agnostic) |
| Candidates | union of 40 dense + 40 lexical candidates per court and the case-number pin (at most 300 passages), with a single joint rerank |
| Context passed | up to 10 decisions (at most 60% from one court, at least 2 from each selected court if relevance ≥ 0.5), for each decision its top 3 reranked passages (not the full text), plus a summary of their decision chains |
| Link extraction | from the operative part and header of the higher decision (regex-based case-number recognition), cross-validated against the repository's related-decision metadata |
| Chain building | the backbone of a chain must be text-confirmed; metadata-only links cannot merge two confirmed chains; the 2013 renaming of Hungarian courts is normalised |
| CJEU link | preliminary-reference cases matched to the underlying Hungarian proceedings via the official register |
| Survival status | deterministic outcome analysis of the higher decision's operative part (upheld / quashed / annulled / complaint rejected), with a downward-propagating safety rule; long chains additionally LLM-audited |
| Ordering | chronological + topological (the appeal relation overrides uncertain dates) |
| Quality | chronology check on every link; ~1% of links are provably wrong → marked uncertain |
| Language model | OpenAI GPT 5.6 Luna (via OpenRouter); fallback: Anthropic Claude Sonnet 5 |
| Streaming | server-sent events (SSE), token-level stream |
| Quote check | verbatim match against the source passages given to the model (after normalising case, typographic marks and whitespace); non-matching quotes are dropped |
| Quote translation | a separate layer AFTER the verbatim check: when a quote's language differs from the interface (e.g. a Hungarian or French quote here, an English CJEU quote on the Hungarian interface), it is shown in machine translation, marked with a “machine translation” badge, and the original is one click away. The checked, original text is never lost — the translation only changes what is displayed. For CJEU quotes, where an authentic translation exists, we use that instead (“official translation”) |
| Backend | FastAPI (Python) |
| Frontend | Next.js (React), Tailwind |
| Authentication | Clerk (JWT) |
| Machine access | MCP (Streamable HTTP) at /mcp as a separate service; OAuth 2.1 via Clerk (dynamic client registration, PKCE); per-user permission and daily allowance |
| Cost tracking | based on actual token usage, per user (admin panel) |
System status
Live availability monitoring for this service: stats.uptimerobot.com/0oGnqSNGvf
This interface is an informational research tool. The content shown does not constitute legal advice. Quotation authenticity is verified mechanically, but always check the original source in any concrete matter.