Tech Stack

Updated Apr 9, 2026

[!info] Compiled from

raw/codebase-snapshots/monorepo-packages-2026-04-09.md

raw/codebase-snapshots/supabase-live-stats-2026-04-09.md

raw/codebase-snapshots/brain-config.ts

Lexery - Technology Stack

Повний реєстр технологій, що використовуються у Lexery. Кожен вибір має причину — дешевий AI через Lexery - Provider Topology|OpenRouter, durable queues через Redis, multi-tenant через Supabase RLS.

Runtime Stack

Layer	Technology	Version / Detail	Purpose
Frontend	Next.js + React 19	`@lexery/portal`	Portal UI — чат, system prompt editor, subscription plans
UI Kit	shadcn/ui + Tailwind CSS	Latest	Component library + utility-first CSS
Backend API	NestJS + Prisma	`@lexery/api`	REST API, auth (email/sms/oauth), storage presigned URLs
Brain Pipeline	Node.js + TypeScript	`@lexery/brain`, port 3081	U1-U12 legal agent runtime, 94 test files
Queue	Redis + BullMQ	`inmemory` or `redis` driver	Event-driven pipeline orchestration, DLQ, reclaim
Run Context	Redis	Compression > 8 KB	Per-run hot state during pipeline processing
Package Manager	pnpm	Workspaces	Monorepo dependency management
Build System	Turborepo	Turbo	Monorepo build orchestration

AI / LLM Layer

Component	Model	Provider	Timeout	Max Tokens	Config Key
Legal reasoning (U10)	`gpt-5.2`	OpenRouter	55s	4,096	`LEGAL_AGENT_MODEL_ID`
Prompt composer (complex)	`gpt-5.2`	OpenRouter	15s	512	`PROMPT_COMPOSER_MODEL_COMPLEX_ID`
Prompt composer (simple)	`gpt-5-nano`	OpenRouter	15s	512	`PROMPT_COMPOSER_MODEL_SIMPLE_ID`
Classification (U2)	`gpt-4o-mini`	OpenRouter	10s	—	`CLF_MODEL_ID`
AI domain classifier	`gpt-4o-mini`	OpenRouter	6s	224	`U2_AI_DOMAIN_MODEL`
Retrieval planner (U4)	`gpt-4o-mini`	OpenRouter	8s	512	`U4_PLANNER_MODEL_ID`
Act planner (U4)	`gpt-4o-mini`	OpenRouter	8s	320-550	`U4_ACT_PLANNER_MODEL`
Query rewrite (U4)	`gpt-4o-mini`	OpenRouter	15s	720	`U4_QUERY_REWRITE_MODEL`
Routing hints (U4)	`gpt-4o-mini`	OpenRouter	10s	384	`U4_ROUTING_HINTS_MODEL`
Repair (U10 post-write)	`gpt-4o-mini`	OpenRouter	—	—	`U10_REPAIR_MODEL_ID`
Embeddings	`text-embedding-3-small`	OpenRouter→OpenAI	5s	1536d	`LLDBI_EMBED_MODEL_ID`

Стратегія: GPT-5.2 тільки для legal reasoning і складних prompt composer calls. Все інше — gpt-4o-mini (дешевий, швидкий, достатньо якісний для structured JSON). Single API key: OPENROUTER_API_KEY_BRAIN > OPENROUTER_API_KEY_ONLINE.

Storage Layer

What	Technology	Provider	Scale
Legal Agent DB	PostgreSQL (Supabase)	Supabase Project 1	370 MB runs, 242 tenants, 935 sessions
Legislation RAG DB	PostgreSQL (Supabase)	Supabase Project 2	374 acts, 42-column schema
Vector DB (legislation)	Qdrant Cloud	Qdrant	`lexery_legislation_chunks` + `lexery_legislation_acts`
Vector DB (memory + docs)	Qdrant Cloud	Same or separate cluster	Semantic memory + MM Docs chunks
Object Storage (legislation)	Cloudflare R2	`legislation` bucket	Canonical legislation JSON (parsed from rada.gov.ua)
Object Storage (runs)	Cloudflare R2	`lexery-legal-agent` bucket	Retrieval traces, MM Docs payloads, query overflow
Queue/Cache	Redis	Self-hosted or managed	BullMQ queues + RunContext store

Зв'язок: Supabase stores metadata and durable state; R2 stores large payloads; Qdrant provides semantic search; Redis handles hot orchestration state. Підхід "best tool for each job" замість monolithic DB.

Monorepo Packages

Package	Type	Stack	Description
`@lexery/brain`	App	Node.js, TS, BullMQ	U1-U12 pipeline runtime
`@lexery/api`	App	NestJS, Prisma, Supabase	Backend API, auth, storage
`@lexery/portal`	App	Next.js 14, React 19, shadcn	User-facing web application
`@lexery/lldbi`	CLI	Node.js, Qdrant, R2	Legislation database management
`@lexery/doclist-resolver`	App	Cloudflare Workers	Act disambiguation API
`@lexery/doclist-updater`	App	Node.js	Rada catalog synchronization
`@lexery/contracts`	Package	Zod	Shared schemas between frontend/backend/agent
Build	Turbo	Turborepo	Workspace orchestration

Infrastructure Services

Service	Provider	Billing	Role
Database (2 projects)	Supabase	Free tier + Pro	Persistence, RLS, auth
Vectors	Qdrant Cloud	Managed	Semantic search over legislation + memory
Objects (2 buckets)	Cloudflare R2	Pay-per-use	Canonical docs, traces, overflow
LLM routing	OpenRouter	Pay-per-token	All AI calls, multi-model
DocList API	Cloudflare Workers	Free tier	Act resolver edge function
CI/CD	GitHub Actions	Free tier	Weekly `brain-admin` batch, linting
VCS	GitHub	`lexeryAI/Lexery`	Private monorepo

Key Config Knobs (brain `lib/config.ts`)

Category	Key Knobs	Default
Rate limiting	`RUNS_PER_MINUTE`, `MAX_CONCURRENT_RUNS`	30 rpm, 10 concurrent
U2 concurrency	`U2_WORKER_CONCURRENCY`, `U2_LLM_CONCURRENCY`	10 workers, 4 LLM
U4 retrieval	`LLDBI_TOP_K`, `U4_GOALS_MAX`, `U4_HITS_CAP`	40, 3, 100
U5 gate	`GATE_MIN_HITS_THRESHOLD`, `GATE_MIN_AVG_SCORE`	3 hits, 0.18 score
U9 budget	`U9_MAX_LAW_SNIPPETS`, `U9_MAX_TOTAL_LAW_CHARS`	20 snippets, 30K chars
U10 write	`LEGAL_AGENT_MAX_TOKENS`, `LEGAL_AGENT_TIMEOUT_SEC`	4096 tokens, 55s
Attachments	`ATTACHMENT_INLINE_MAX_BYTES`	512 KB inline, 2 MB request
Query overflow	`QUERY_R2_THRESHOLD_BYTES`	32 KB → R2

Повний config: 635 рядків у apps/brain/lib/config.ts — кожна стадія має свої env knobs із bounded ranges.

Data Sources and Volumes

Source	Volume	What
rada.gov.ua	374 acts (360 in force)	Ukrainian legislation, parsed → Qdrant + R2
Supreme Court case law	~1000 entries in R2	Case law for future RAG expansion
User queries	26,704 runs	Legal questions in Ukrainian
User uploads	679 docs	Documents attached via MM Docs
Memory extraction	3,565 items	Semantic user knowledge
Messages	7,239	User + assistant conversation messages

CI/CD Pipeline

Trigger	What	Details
`brain-admin` (weekly)	LLDBI batch: import proposals, act updates	GitHub Actions, `apps/lldbi/brain-admin`
pnpm test scripts	94 test files covering U2-U12, MM, ORCH, gateway	Local `tsx` runners + CI
Manual smoke	`brain:verify:*`, `brain:concurrency:smoke`	Live dev environment validation