A lawyer receives an argument map with exact citations.
Every result opens the full judgment. When the material offers no basis, the system says so clearly. Interpretation and argument choice remain with the lawyer.
Problem, solution, and result
The problem
Case-law research requires opening many records and comparing their reasoning. Much of the time goes into finding passages that may matter to the case.
How the system works
The system searches an approved set and returns propositions linked to exact source passages. Each citation opens the full judgment. If the material offers no support, the result states that clearly.
What we tested
The demonstration returned citations that matched the stored text and refused questions outside its scope. It does not assess the legal relevance of an argument or performance on the firm’s full corpus. A lawyer must confirm both during a pilot.
For whom
This is a good automation candidate when lawyers repeatedly search the same corpus and spend more time locating passages than evaluating them.
Question → passages → propositions → sources
- 01Every proposition requires an exact citation
- 02The citation opens the full SAOS record
- 03The lawyer interprets and chooses the argument
- Business type
- Law firms and in-house legal teams
- Input
- A lawyer’s research question and 40 SAOS ruling excerpts with source links and retrieval times
- Boundary
- A proposition without an exact citation span fails validation; a request for advice ends in a refusal
- Cost
- PLN 0 in external API fees for the recorded local run covering 10 questions.
- Security
- An answer without exact support in an approved source is held back. The system refuses requests for legal advice.
- Speed
- We measure retrieval time in a pilot using the firm’s corpus and questions.
- Trace
- Each proposition retains its source, exact passage, and any refusal decision.
- Similar system build
- from €6,000 net (approx. $6,500) · 4–10 weeks
Where automation stops
The lawyer interprets the result
The system stops at propositions and citations. The lawyer checks the full reasoning, the current state of the law, and whether an argument fits the matter. A request for legal advice ends in a refusal. The public version replays 10 saved questions and accepts no visitor text.
- Boundary
- A proposition without an exact citation span fails validation; a request for advice ends in a refusal
- Cost
- PLN 0 in external API fees for the recorded local run covering 10 questions.
- Security
- An answer without exact support in an approved source is held back. The system refuses requests for legal advice.
- Speed
- We measure retrieval time in a pilot using the firm’s corpus and questions.
Why we show the lower score
The first setup scored 0.950 because it promoted the topic used to create the labels. Reading all 40 excerpts exposed the issue, and the honest hybrid score was 0.735. We removed that rule. The next run produced 0.870 for hybrid search and 0.894 for BM25. Three of 10 questions had no responsive excerpt, and the system reported that. Legal validation requires a larger lawyer-labeled set.
Estimated impact
Calculate the impact at your volume
This estimate uses the stated volume. Enter your own numbers to assess the possible impact in your company. The result is confirmed only during a pilot.
Today
280 h
With the system
90 h
Time or cost saved
190 h/mo
- Volume
- 600 questions/mo
- Formula
- (600 × (28 - 9)) / 60 = 190 h/mo
- Calculation status
- modelled scenario
Data in the screenshots. The names, amounts, and documents shown in the screenshots are synthetic. Client data remains private. Measurements describe the demonstrator and are kept separate from production outcomes.
Working surfaces
The lawyer evaluates the citation, the current law, and the argument to use in the matter.
The lawyer starts with propositions, citations, and gaps in the material. Competing views stay side by side. Full reasoning still needs legal review. Quality on firm data is measured on a larger question set labeled by lawyers.
Three views for case-law research
Question list
Saved questions and whether the corpus contains relevant material.
Argument map
Propositions, competing views, and exact quotations side by side.
Source record
Full judgment context, official URL, and provenance data.
System screens
See how the system works in practice
These desktop and mobile screenshots come from the running application. They show the process and the points where a person makes the decision.
- Screens
- 16
- px
- 1440 · 390
- 021440×1702
A citation expanded to its context and full source record. - 031440×1100
A clear result when the limited corpus contains no support.
Open the remaining screen archive (13)
- 041440×1100
A split in the corpus - 051440×1177
Both views and the lawyer's gate - 061440×1100
Research questions - 071440×1100
A question with no responsive excerpt - 081440×1100
Cited spans - 091440×1100
Rulings in the corpus - 101440×1100
A selected excerpt and its SAOS record - 111440×1100
Questions sent back to a lawyer - 121440×1100
Saved queries - 131440×1100
Research history - 141440×1100
Measurement method and ranking correction - 151440×1100
Retrieval rules - 161440×1100
Read labels and the review boundary
Technology stack
Retrieval runs locally, and a citation is required
BM25, multilingual-e5-small, and a reranker retrieve passages without an external API. CitedProposition requires at least one CitationSpan.
- Next.js
- a read-only public view with the analysis and cited passages
- FastAPI
- retrieval and the response contract with its required CitationSpan
- BM25 + multilingual-e5-small + cross-encoder
- hybrid ranking of excerpts, fully offline
- PostgreSQL
- documents, indexes, and the run trace
- Docker Compose
- API, database, and retrieval on the internal network; only the replay faces outward
The client receives the code, index, test data, and documentation. Production use requires an approved source corpus and legal evaluation.
Technical details and measurements
The working loop
From question to a passage that can be checked
The system searches, ranks, and validates citation spans. A proposition without a matching source passage cannot appear in the result.
Research question
Search in an approved corpus
Propositions with exact citations
No support → clear statement
Lawyer checks the full judgment
System architecture
Frozen corpus, local retrieval, and citation validation
PostgreSQL stores public records. Local models retrieve passages, and the data contract requires an exact citation span for every proposition.
- 01
Corpus
A frozen set of rulings with proof of origin.
Every excerpt enters the corpus with its URL, retrieval time, and SHA-256 content hash, so every answer can be traced back to its source.
- 02
Retrieval
BM25, multilingual-e5-small, and a cross-encoder, all local.
Search combines lexical and semantic matching with reranking, because a legal question can be a paraphrase and a case signature is a string of characters. With topic promotion removed, hybrid scores 0.870 and BM25 0.894 against read labels. Queries never leave the environment.
- 03
Citation
CitedProposition requires a non-empty CitationSpan.
A proposition without an exact span in the source text fails schema validation and never reaches the interface. The no-invented-sources guarantee lives in the data type, so it does not depend on prompt discipline.
- 04
Refusal and audit
Research has an explicit end.
A request for advice, document drafting, missing material, or sensitive data ends in a refusal. The public screen is read-only. The API, database, and search stay on the internal network.
One process rather than a conversation
The system holds no conversation and improvises nothing. It performs one process: find the excerpts, build propositions with spans, show the sources, and refuse where research ends. The simpler design is deliberate: there is no agent loop and no generative summarising here, because the citation guarantee comes from a simpler mechanism, a data contract checked before anything renders.
- A proposition is formed only from a passage present in the verified corpus
- Documents and queries stay in the local environment
- Requests for legal advice end in an explicit refusal
- The current run's metrics and limitations are visible in the interface
Want to assess a similar process in your company?
- A 30-minute call with the engineer who would lead the work.
- A review of the processes that cost you the most time and money.
- A written summary of what to automate first and the likely cost range.
Times are shown in your own time zone. We work with clients across time zones.
The scan is free and creates no obligation. If automation is unlikely to pay off, the written recommendation will say so.