Lexery Wiki

Lexery - Brain Architecture

4 хв читання

Compiled from

  • raw/architecture-docs/LEXERY_LEGAL_AI_AGENT_ARCHITECTURE.md
  • raw/architecture-docs/app-README.md
  • raw/architecture-docs/MEGA_DIAGRAM_FULL.md

Lexery - Brain Architecture

Best Short Definition

apps/brain is the current core legal runtime of Lexery: a staged, evidence-first, bounded legal agent rather than a single-prompt chatbot.

What Brain Owns

Observed boundaries

  • Legal runtime and stage orchestration
  • Verification tooling
  • Run-level state handling
  • Prompt assembly across law, docs, memory, history
  • Writer / verifier / deliver path

Explicit non-ownership

  • LLDBI corpus/admin infra lives outside Brain in apps/lldbi
  • DocList services live in apps/doclist-*
  • Product shell/control plane lives in apps/api and apps/portal

Current Module Atlas

  • gateway/
  • classify/
  • plan/
  • retrieval/
  • gate/
  • expand/
  • assemble/
  • reasoning/
  • write/
  • verify/
  • deliver/
  • mm/
  • orchestrator/
  • tools/
  • lib/

Current Runtime Shape

Earlier graph

U1 -> U2 -> U3 -> U3a -> U4 -> U5 -> (U6|U9) -> U10 -> U11 -> U12

Upgraded graph

U1 -> U2 -> (U3 | ORCH) -> U3a -> U4 -> U5 -> ORCH -> (U6 | U7) -> ORCH -> (U4 | U8 | ASK_USER) -> ORCH -> U9 -> ORCH -> U10 -> ORCH -> U11 -> ORCH -> U12

Meaning

  • Observed: Brain moved from a mostly linear staged pipeline to a hybrid staged-plus-orchestrated runtime.
  • Observed: deterministic fast paths still exist.
  • Inferred: the project is trying to gain agentivity without losing legal controllability.

Important Architectural Principles

Principle 1 — retrieval-first legal grounding

  • U4 is treated as the strongest intelligence layer for law retrieval.
  • Writer is downstream of evidence, not a free generator.

Principle 2 — bounded orchestration

  • ORCH appears only where the next move is no longer single-path.
  • Not every stage transition is handed to an LLM.

Principle 3 — explicit state surfaces

  • runs row
  • runs.snapshot
  • RunContext
  • retrieval/gate/orch/clarification substructures
  • R2 overflow storage

Principle 4 — additive compatibility

  • Existing API and run semantics stay stable where possible.
  • New orchestration state is added into snapshot/context rather than by hard schema explosion.

Transitional Reality

Observed

  • write/verifyConsumer.ts and write/deliverConsumer.ts are still the live U11/U12 entrypoints.
  • verify/ and deliver/ directories remain placeholders.
  • expand/consumer.ts is the live U6 entrypoint.

Meaning

  • The architecture is real, but module boundaries are still being cleaned up.

Quality Posture

  • stage-specific verifiers are important
  • retrieval verifiers intentionally isolate upstream truth from downstream writer noise
  • real-mode runs are cost-bounded
  • docs emphasize honest degradation, not hallucinated completeness

Best Reading Of Brain Maturity

  • Observed: Brain is the most technically mature domain surface in the project.
  • Observed: its stage vocabulary is stable enough to act as the project’s shared ontology.
  • Inferred: a lot of Lexery’s strategic moat, if it works, will live here rather than in generic chat UI.

Key Sources

  • apps/brain/README.md
  • apps/brain/LEGAL_AGENT_README.md
  • docs/lexery-current-runtime-map.md
  • apps/brain/docs/architecture/app/**

Monorepo Structure

PackagePathPurpose
@lexery/brainapps/brainLegal Agent Brain: HTTP server, U1–U12 pipeline, retrieval, MM / MM Docs
@lexery/lldbiapps/lldbiLLDBI admin CLI: Qdrant / R2 / Supabase legislation ops
@lexery/apiapps/apiNestJS backend (Prisma, Supabase, S3)
@lexery/portalapps/portalNext.js frontend (React 19)
@lexery/doclist-updater-dbapps/doclist-updater-dbDaily DocListDB incremental updater (Rada → embeddings → Qdrant)
@lexery/doclist-full-importapps/doclist-full-importFull DocList catalog importer into Qdrant
@lexery/doclist-resolver-apiapps/doclist-resolver-apiCloudflare Worker: act resolution via Qdrant
@lexery/typescript-configpackages/config/typescriptShared TypeScript config

Architecture Docs in Repo

Canonical architecture documentation lives under apps/brain/docs/architecture/ (100+ markdown files), including:

  • LEXERY_LEGAL_AI_AGENT_ARCHITECTURE.md — full-stack narrative (Ukrainian)
  • MEGA_DIAGRAM_FULL.md — system-wide Mermaid diagram
  • app/README.md — implementation index mapping U1–U12 to consumers
  • app/context/CURRENT_PIPELINE_STATE.md — live snapshot of what runs (updated 2026-04-09)

Testing Infrastructure

Roughly 61 test-oriented TypeScript modules under apps/brain/tools/, with naming conventions:

  • test_*_units.ts — unit test suites per stage
  • stress_test_*.ts — load / stress tests (especially U4 retrieval)
  • verify_*.ts — live verification harnesses

Script aliases in the monorepo: pnpm brain:test:* / pnpm brain:verify:*.

CI/CD

Single GitHub Action: .github/workflows/lldbi-brain-admin.yml

  • Weekly cron: Monday 04:15 UTC plus manual workflow_dispatch
  • Flow: brain-admin batch → process approved proposals → upload artifacts (report.json, approved-proposals-report.json)

See Also

Вигляд графа

Зворотні посилання