Glossary

Glossary

Essential terms from the Brazilian legal domain and Valter’s technical architecture. Understanding these terms is critical for working with the codebase and reading the documentation.

Legal Domain

A

Acórdão A collegial (collective) court decision, issued by a panel of ministers rather than a single judge. Most STJ decisions indexed by Valter are acórdãos.

C

Critério A legal criterion or standard applied by the court to decide a case. In the knowledge graph, criteria are nodes that connect multiple decisions applying the same legal test.

D

Desembargador A justice of a regional court (TRF). Relevant for multi-tribunal expansion in v2.0. Equivalent to “judge” at the appellate level.

Dispositivo Legal A legal provision — a specific article, paragraph, or clause of a statute or regulation cited in a court decision. In the knowledge graph, dispositivos connect decisions that apply the same legal text.

Divergência A disagreement between ministers or turmas on how to apply a legal criterion. Valter’s graph analytics can detect and track active divergences, which are valuable for identifying unsettled areas of law.

E

Ementa The official summary or headnote of a court decision. This is the most commonly indexed text in Valter’s corpus. Most of the ~23,400 documents have at least an ementa; fewer have the full text (íntegra).

F

Fase A procedural phase or stage in a case (e.g., initial petition, defense, judgment). Used in the PROJUDI pipeline for classifying document types within a case file.

I

Íntegra The full text of a court decision, as opposed to just the ementa (summary). Downloading and indexing íntegras improves search quality and enables deeper analysis.

IRAC Issue, Rule, Application, Conclusion — a standard legal analysis framework. Valter uses IRAC to structure the analysis of court decisions, breaking each into its constituent reasoning steps. Implemented in models/irac.py.

J

Jurisprudência The body of court decisions on a topic. “Case law” in English. Valter’s primary domain is STJ jurisprudência on federal law matters.

M

Ministro A justice of a superior court (STJ, STF). Called “minister” in Portuguese court terminology. Valter tracks minister voting patterns, specializations, and divergences.

Ministro Relator The reporting minister — the justice assigned to a case who writes the decision and presents it to the panel. The relator’s identity is a key metadata field for every indexed decision.

N

Número de Registro The registration number of a decision in the court’s internal system. Used as a unique identifier when linking to STJ public records.

P

Precedente A prior court decision used as the basis for a legal argument. In the knowledge graph, precedent relationships form citation chains that the reasoning engine traverses.

PROJUDI (Processo Judicial Digital) The electronic judicial process system used in Paraná (PR) state courts. Valter has a specific pipeline (core/projudi_pipeline.py) for extracting and classifying documents from PROJUDI case files.

R

REsp (Recurso Especial) A special appeal to the STJ on matters of federal law interpretation. REsps are the most common type of decision in Valter’s corpus.

S

Seção A section of the court, composed of two turmas. The STJ has three sections: 1st Section (1st + 2nd Turma), 2nd Section (3rd + 4th Turma), and 3rd Section (5th + 6th Turma). Sections hear cases when turmas disagree.

seqDocumento The sequential document identifier in STJ’s internal system. Used for linking directly to specific decisions on the STJ portal.

STF (Supremo Tribunal Federal) Brazil’s Supreme Federal Court. Handles constitutional matters. Planned for v2.0 multi-tribunal expansion.

STJ (Superior Tribunal de Justiça) Brazil’s Superior Court of Justice. Responsible for unifying the interpretation of federal law. Valter’s primary and currently only data source.

Súmula A binding legal summary issued by a superior court (STJ or STF) that consolidates the court’s understanding on a specific legal question. Súmulas are numbered (e.g., Súmula 123/STJ) and carry significant precedential weight.

T

TRF (Tribunal Regional Federal) Federal Regional Courts at the appellate level. Brazil has six TRFs. Planned as the first expansion target in v2.0.

TST (Tribunal Superior do Trabalho) The Superior Labor Court. Handles labor law appeals. Planned for v2.0.

Turma A panel or division of ministers that hears cases. The STJ has six turmas. Each minister belongs to one turma. Turma assignment affects which cases a minister hears and can reveal patterns in voting behavior.

---

Technical Terms

A

ARQ (Async Redis Queue) A Python library for background job processing using Redis as a broker. Valter uses ARQ for asynchronous ingestion tasks in workers/.

B

BM25 (Best Matching 25) A probabilistic text ranking algorithm used for lexical (keyword-based) search. One of two retrieval strategies in Valter’s hybrid search, alongside semantic vector search.

C

Canary Rollout A deployment strategy where a small percentage of traffic is gradually routed to a new system to validate it before full migration. Applied to R2 artifact storage in stores/artifact_storage.py.

Circuit Breaker A resilience pattern that stops calling a failing service after repeated failures, allowing it to recover. Planned for v1.1 to prevent Neo4j hangs from blocking all requests.

Cross-Encoder A neural model that jointly encodes a query and a document together for fine-grained relevance scoring. Used in the reranking stage of search, after initial retrieval by BM25 and semantic search.

D

Dual-Vector An encoding strategy that creates separate embeddings for facts and legal thesis within a document, enabling more targeted retrieval. Implemented in core/dual_vector_retriever.py.

F

FRBR (Functional Requirements for Bibliographic Records) A conceptual model for bibliographic records, adapted as the ontology for Valter’s Neo4j knowledge graph. Structures how decisions, provisions, and concepts relate to each other.

K

KG Boost (Knowledge Graph Boost) A relevance boost applied to search results based on their connectivity in the Neo4j knowledge graph. Documents with more citations, shared criteria, and graph connections receive higher scores. Configurable via VALTER_KG_BOOST_BATCH_ENABLED.

M

MCP (Model Context Protocol) An open standard for LLM-to-tool integration. Allows language models like Claude and ChatGPT to use external tools through structured tool definitions with JSON Schema. Valter supports two MCP transports: stdio (local, for Claude Desktop) and HTTP/SSE (remote, for ChatGPT and other consumers).

R

RRF (Reciprocal Rank Fusion) A method for combining ranked lists from different retrieval strategies into a single ranking. Valter uses RRF to merge BM25 lexical results with semantic vector results, producing a hybrid ranked list.

S

structlog A structured logging library for Python. Valter uses structlog for JSON-formatted logs with request-scoped trace_id for tracing requests across the system.