01
Classification and structuring of large case law repositories
A firm accumulates hundreds of decisions on the same subject with no shared structure: each one is read in full every time the question comes up again.
- Definition of a schema of legally relevant fields (court, date, question at issue, position taken, outcome).
- Assisted extraction and normalisation of those fields across the whole corpus.
- Human review by sampling and correction of the schema before the corpus is treated as valid.
A repository searchable by legal position rather than by free text alone, allowing the applicable line of authority to be located in minutes and divergent decisions to be spotted.
02
Systematic verification of case law citations
Submissions inherit citations from earlier templates. A single mistranscribed, superseded or non-existent reference is enough to undermine the credibility of the whole document.
- Automatic extraction of every legislative and case law reference in a document.
- Checking of each reference against the primary source (official gazette, court database, tax tribunal doctrine).
- A verification report stating the status of each citation: confirmed, corrected, not located or superseded.
A reproducible quality control step before filing, with a record of what was verified and against which source.
03
RAG system over validated legal documentation
General-purpose models answer fluently but without a verifiable source, which makes them unusable for work that must be capable of being cited.
- Definition of a closed, validated corpus: consolidated legislation, case law and administrative criteria.
- Indexing with passage-level retrieval and answers necessarily anchored to documents in the corpus.
- System design so that every statement returns its source passage and the absence of a source is declared as such.
Internal queries answered with verifiable references and an explicit limit: if the corpus does not support it, the system does not assert it.
04
Automation of research matrices
Research on a recurring question is redone from scratch for every matter, and the result depends on who carries it out.
- Definition of the matrix: questions to be answered, mandatory sources and acceptance criteria.
- Assisted population of the matrix for each new matter, with empty fields flagged explicitly.
- Versioning of the matrix so that the position and its evolution are preserved over time.
A consistent, auditable and transferable research method within the team, regardless of who executes it.