Skip to content

Roadmap

Valter’s evolution from legal search backend to legal reasoning engine — strategic vision, competitive positioning, and risk-informed planning.

Valter will be the legal reasoning engine for LLMs and lawyers. The core differentiator is not search — it is the ability to compose verified legal arguments from a knowledge graph containing ~28,000 STJ decisions and 207,000 relations.

The platform combines six capabilities that do not exist together in any other legal tech product:

  1. Anti-hallucination verification against real tribunal data
  2. Temporal intelligence — weighting by recency, trend detection across jurisprudence
  3. Judge profiling — delta analysis by minister, active divergence tracking
  4. Knowledge graph — decisions connected by criteria, legal provisions, and precedents
  5. MCP-native architecture — any LLM (Claude, ChatGPT) can use Valter as a tool
  6. Argument composition — assembling multi-step legal reasoning chains from graph patterns

The v1.x series stabilizes production, adds resilience, and launches the Legal Reasoning Chain (the flagship feature). The v2.x series expands to multiple tribunals, integrates sister products (Leci, Juca), and establishes public presence via the ChatGPT App Directory.

DimensionOthersValter
SearchKeyword searchHybrid search with KG boost and cross-encoder reranking
ResultsList of similar casesArgument composition from success patterns in graph
Citations”The jurisprudence says…""REsp X says Y, verified, cited Z times in graph”
WeightingTreat all decisions equallyTemporal intelligence — weights by recency, trend detection
JudgesIgnore who judgesJudge profiling — delta by minister, active divergences
AccuracyHallucinate referencesAnti-hallucination verification against real tribunal data
IntegrationClosed systemsMCP-native — any LLM can use as tool
Data modelDocuments as text blobsKnowledge graph — decisions connected by criteria, provisions, precedents

As of February 2026, Valter has:

  • 33 implemented features spanning search, graph analytics, MCP server, ingestion, verification, observability, and CI/CD
  • 3 features in progress: App Directory preparation (~70%), R2 canary storage (~90%), missing integras download flow (~80%)
  • 18 planned features distributed across v1.0 through v1.2
  • 11 ideas under evaluation for v2.0+

The system indexes ~23,400 STJ decisions in PostgreSQL, with ~3,700 vectorized in Qdrant and ~28,000 nodes with ~207,000 relationships in Neo4j.

Eight failure scenarios were analyzed to identify risks before they materialize. Each scenario maps to a specific milestone that addresses it.

#Failure ScenarioFalse PremiseMitigationMilestone
1Redis goes down, all requests blocked”Redis will always be up”Fail-open rate limiter for valid API keysv1.0
2Neo4j silently returns stale data”Graph data is always fresh”Health check endpoint, staleness alertsv1.0
3Data stagnation — corpus stops growing”23K docs are enough”ARQ cron ingestion, indexation gap closurev1.0 / v1.1
4Neo4j query hangs, blocks entire request”Neo4j always responds in time”Circuit breaker with 5s timeoutv1.1
5Multi-tribunal is much harder than expected”Just add more data”TRF spike in v1.2 before committing to v2.0v1.2
6App Directory yields low ROI”ChatGPT users will find us”Demoted from v1.2 to v2.1, focus on core productv2.1
7Single developer becomes unavailable”Diego can always fix things”Absence runbook, documented ops proceduresv1.0
8Docker compose issues in new environments”It works on my machine”Documented setup, pinned versions, CI validationv1.0

Nine architectural and product decisions remain open. Each will be resolved before or during its target milestone.

#DecisionOptions Under ConsiderationTarget
1R2 canary activation timingActivate in v1.0 vs wait for more traffic datav1.0
2Legacy route sunsetImmediate removal vs deprecation period with warningsv1.0
3Privacy policy / terms of use authorshipWrite in-house vs engage legal counselv1.0
4Leci integration modelEmbedded library vs API calls vs shared databasev2.0
5Juca integration levelRead-only consumer vs bidirectional syncv2.0
6Multi-tribunal starting pointTRF-4 (most data) vs TST (simpler schema) vs STF (highest impact)v2.0
7Doutrina (legal doctrine) scopeIndex doctrine texts vs link to external sources onlyv2.0+
8Embedding model migrationKeep Legal-BERTimbau 768d vs upgrade to 1024d modelv1.1
9Reasoning chain sync vs asyncSynchronous endpoint vs background job with pollingv1.2