73 lines
3.0 KiB
Markdown
73 lines
3.0 KiB
Markdown
# Implementation Plan: Unified AI Architecture
|
|
|
|
**Branch**: `301-unified-ai-arch` | **Date**: 2026-05-14 | **Spec**: [spec.md](./spec.md)
|
|
**Input**: Feature specification from `specs/300-others/301-unified-ai-arch/spec.md`
|
|
|
|
## Summary
|
|
|
|
Implement a Master AI Architecture enforcing strict physical isolation of AI workloads on a dedicated Admin Desktop (Desk-5439). The system supports secure legacy document migration via a staging queue, context-aware conversational RAG queries, and detailed AI audit logging, all orchestrated through robust backend queues (BullMQ) and multi-tenant security filters.
|
|
|
|
## Technical Context
|
|
|
|
**Language/Version**: TypeScript (Node.js v24.15.0) for Backend/Frontend
|
|
**Primary Dependencies**: NestJS 11, Next.js 16, BullMQ, Qdrant Node Client, n8n
|
|
**Storage**: MariaDB 11.8 (Relational), Qdrant (Vector), Redis (Queue/Cache)
|
|
**Testing**: Jest (Backend), Vitest (Frontend), E2E with Playwright
|
|
**Target Platform**: QNAP Container Station (Production), Desk-5439 (AI Host)
|
|
**Project Type**: Monorepo Web Application (Backend + Frontend)
|
|
**Performance Goals**: RAG Response < 10s (p95), Migration throughput 1000 pages/hour
|
|
**Constraints**: AI host has limited VRAM (8GB), necessitating concurrency limit of 1 for LLM generation.
|
|
**Scale/Scope**: 20,000+ legacy documents, project-wide deployment.
|
|
|
|
## Constitution Check
|
|
|
|
_GATE: Must pass before Phase 0 research. Re-check after Phase 1 design._
|
|
|
|
- **ADR-019 UUID**: `publicId` used exclusively. No INT primary keys exposed.
|
|
- **ADR-009 Database**: Schema changes via raw SQL deltas.
|
|
- **ADR-016 Security**: CASL RBAC strictly enforced (`@UseGuards(CaslAbilityGuard)`). Idempotency-Key headers required.
|
|
- **ADR-008 BullMQ**: Heavy AI orchestration and RAG queuing managed via BullMQ.
|
|
- **ADR-018/023 AI Boundary**: AI host connects via DMS API. No direct database access.
|
|
- **TypeScript Strict**: Explicit types, no `any`, proper error handling via `BusinessException`.
|
|
|
|
## Project Structure
|
|
|
|
### Documentation (this feature)
|
|
|
|
```text
|
|
specs/300-others/301-unified-ai-arch/
|
|
├── spec.md
|
|
├── plan.md # This file
|
|
├── research.md
|
|
├── data-model.md
|
|
├── quickstart.md
|
|
├── contracts/
|
|
└── tasks.md # (To be created)
|
|
```
|
|
|
|
### Source Code (repository root)
|
|
|
|
```text
|
|
backend/
|
|
├── src/
|
|
│ ├── ai/
|
|
│ │ ├── ai.module.ts
|
|
│ │ ├── ai.controller.ts
|
|
│ │ ├── ai.service.ts
|
|
│ │ ├── qdrant.service.ts
|
|
│ │ ├── rag.processor.ts
|
|
│ │ └── dto/
|
|
│ └── database/
|
|
│ └── sql/
|
|
|
|
frontend/
|
|
├── src/
|
|
│ ├── app/(dashboard)/ai-staging/
|
|
│ ├── components/ai/
|
|
│ │ ├── AiStatusBanner.tsx
|
|
│ │ └── RagChatWidget.tsx
|
|
│ └── lib/api/ai.ts
|
|
```
|
|
|
|
**Structure Decision**: Integrated into the existing Next.js / NestJS monorepo architecture, utilizing a dedicated `AiModule` in the backend to centralize all external AI API calls and queue management.
|