Skip to content
BranchesTags

Repository files navigation

Enterprise AI Chat

Enterprise-grade AI chat platform with multi-provider support, intelligent model orchestration, autonomous AI agents, document studio, marketplace, and VS Code extension.

Current version: 2.1.65

Repository: github.com/marypas74/ai_enterprise

Features

  • Intelligent Model Orchestrator — Automatic model selection via rule-based scoring + semantic embedding routing. Routes queries to Fast/Balanced/Powerful tiers with circuit breaker, cascade escalation, and feedback loop.
  • Multi-Provider AI — OpenAI, Anthropic Claude, Google Gemini, Ollama (local models) with automatic failover. LiteLLM proxy support.
  • Autonomous AI Agents — Claude Agent SDK integration with terminal orchestration, worktree isolation, task management, and iterative execution via Auto-Claude dashboard.
  • Document Studio — PDF editor (pdftohtml + LibreOffice + Ollama Vision OCR), PAdES-B-B digital signatures, DOCX/XLSX/PPTX generation, OnlyOffice integration.
  • Marketplace — Plugin/skill catalog with approval workflow, catalog service, KB integration, Qdrant vector search, and backend client for automated installs.
  • Vector Memory — 4-tier RAG pipeline (episodic/declarative/procedural/working) with HyDE, embeddings, semantic search, and memory stats dashboard.
  • Async Document Queue — Large RAG requests (>8K tokens) are intercepted and dispatched to a background DocumentJobWorker with Redis-backed job queue, WebSocket notifications (/ws/jobs), and stale job recovery on worker restart.
  • Summary Intent Detection — Document queries are classified as "summary" vs "specific" to either pull distributed chunks covering the whole document or run semantic search (IT/EN patterns for "riassumi", "di che argomenti/temi/tematiche parla", "what is this document about", etc.).
  • Plugin System — File-based plugins with EventBus hooks, MCP server support, skill management, prompt templates.
  • EU AI Act Compliance — Art. 50.1/50.2 disclosure, consent management, bias monitoring, audit logging, AI transparency page, DPIA documentation.
  • VS Code Extension — 19 commands: chat, code explain/fix/improve/document, agent sessions, inline editing, provider switching.
  • Voice Interface — OpenAI Whisper STT + OpenAI TTS / local Piper TTS fallback with animated Avatar Orb overlay.
  • Image Generation — OllamaDiffuser integration (FLUX.1 schnell) inline in chat.
  • Admin Dashboard — User/group management, provider config, orchestrator dashboard, pipeline visualizer, plugin graph, hooks, guides, kanban.
  • Security — JWT + MFA (TOTP), Google OAuth, Zod input validation, OWASP hardening, rate limiting, network policies.
  • Kubernetes Ready — Production deployment on MicroK8s with auto-scaling, backup CronJobs, Cloudflare Tunnel, OnlyOffice, Open-WebUI, Parlant.
  • Mobile — Android APK (Capacitor 6) with native SSE + PWA for iOS (installable from Safari).

Architecture

                     ┌──────────────────────────────────────────────────────┐
                     │              Frontend (React 18 + Vite)              │
                     │  Chat │ Admin │ Projects/Kanban │ Agents │ Documents │
                     │  Zustand stores │ Tailwind CSS │ Playwright E2E      │
                     └───────────────────────┬──────────────────────────────┘
                                             │ Nginx reverse proxy
┌────────────────────────────────────────────▼──────────────────────────────────┐
│                      Backend (Fastify 5 + TypeScript)                         │
│  25 modules │ JWT/MFA/OAuth │ WebSocket │ SSE streaming │ Zod validation      │
├──────────────────────────────────────────────────────────────────────────────┤
│                         Model Orchestrator                                    │
│                                                                               │
│  User Query ──▶ ModelRouter ──▶ Tier Selection ──▶ Provider                   │
│                    │                │                  │                      │
│              Rule-based        Semantic          Circuit Breaker              │
│              scoring          embedding           + Cascade Fallback          │
│                    ▼                ▼                   ▼                     │
│         ┌──────────────┐  ┌──────────────┐  ┌──────────────────┐            │
│         │    FAST      │  │   BALANCED   │  │    POWERFUL      │            │
│         │ Haiku 4.5    │  │ Sonnet 4.6   │  │ Opus 4.6         │            │
│         │ GPT-4.1-mini │  │ GPT-4.1      │  │ o3-mini          │            │
│         │ Gemini Flash │  │ Gemini Pro   │  │                  │            │
│         │ Ollama       │  │              │  │                  │            │
│         └──────────────┘  └──────────────┘  └──────────────────┘            │
├────────┬──────────┬────────┬──────────┬──────────┬──────────┬───────────────┤
│  Auth  │   Chat   │ Agents │ Projects │  Tools   │ Docs     │ Compliance    │
│  MFA   │ Complete │ SDK    │  Kanban  │ PDF/DOCX │ Studio   │ EU AI Act     │
│  OAuth │ Stream   │ Orchst │  Boards  │ PPTX/XLS │ OnlyOff  │ Consent/Audit │
│  Sessi │ Memory   │ Worktr │  Cards   │ Sandbox  │ PAdES-BB │ Bias Monitor  │
└────────┴────┬─────┴────────┴──────────┴──────────┴──────────┴───────────────┘
              │
 ┌────────────┼────────────┬──────────────┬──────────────┬────────────────┐
 │            │            │              │              │                │
┌▼───────┐ ┌──▼─────┐ ┌────▼─────┐ ┌─────▼────┐ ┌──────▼──────┐ ┌──────▼──┐
│MariaDB │ │ Redis  │ │  Qdrant  │ │ Parlant  │ │   Ollama    │ │LiteLLM  │
│Users   │ │Session │ │ Vectors  │ │ Agents   │ │ Local LLMs  │ │ Proxy   │
│Chat    │ │Cache   │ │Embeddings│ │Guidlines │ │ GPU/Docker  │ │ Router  │
│History │ │Tokens  │ │ RAG/KB   │ │ Sessions │ │             │ │         │
└────────┘ └────────┘ └──────────┘ └──────────┘ └─────────────┘ └─────────┘

Model Orchestrator

Automatically selects the optimal AI model for each query without user intervention.

How It Works

  1. Rule-Based Router — Analyzes query length, keywords, attachments, conversation depth, and tool usage to compute a complexity score.
  2. Semantic Router — Uses embedding similarity against pre-computed route examples for sub-millisecond task classification (greeting, coding, analysis, complex reasoning, etc.).
  3. Response Quality Checker — Evaluates response quality without LLM calls (refusal detection, truncation, uncertainty). Supports cascade escalation to a more powerful model.
  4. Circuit Breaker — Tracks provider health, opens circuit on consecutive failures, auto-recovers.
  5. Feedback Loop — Records routing decisions with latency, cost, and user overrides. Admin dashboard shows routing distribution and cost savings.

Routing Tiers

Tier Default Models Use Case
Fast Haiku 4.5, GPT-4.1-mini, Gemini Flash, Ollama Greetings, simple questions, formatting
Balanced Sonnet 4.6, GPT-4.1, Gemini Pro Coding, analysis, writing, standard tasks
Powerful Opus 4.6, o3-mini Architecture, complex reasoning, multi-step agents

Quick Start

Prerequisites

  • Node.js 20+
  • Docker
  • MicroK8s (Kubernetes deployment)

Local Development

# Clone
git clone https://github.com/marypas74/ai_enterprise.git
cd ai_enterprise

# Backend
cd backend
npm install
cp .env.example .env    # Configure API keys, DB, Redis, JWT secrets
npm run dev             # Hot reload on port 3000

# Frontend
cd frontend
npm install
npm run dev             # Vite dev server on port 5173 (proxies /api → :3000)

# Tests
cd backend && npm test              # Vitest unit tests
cd frontend && npm test             # Vitest + Testing Library
cd frontend && npm run test:e2e     # Playwright E2E

# VS Code Extension
cd vscode-extension
npm run build:all   # Build webview + compile extension
npm run package     # Create .vsix

Kubernetes Deployment

bash BUILD.sh       # Full build: npm install + Docker + MicroK8s import + K8s deploy
sudo bash DEPLOY.sh # Quick deploy: import pre-built images + restart pods

Project Structure

ai_enterprise/
├── backend/                    # Fastify 5 API server (TypeScript)
│   └── src/
│       ├── modules/            # 25 feature modules
│       │   ├── auth/           # JWT + MFA (TOTP) + Google OAuth
│       │   ├── chat/           # Completions, conversations, models, SSE streaming
│       │   ├── admin/          # Users, providers, plugins, settings, orchestrator, guides
│       │   ├── agents/         # AI agent sessions + Claude Agent SDK routes
│       │   ├── projects/       # Kanban boards, cards, access control
│       │   ├── memory/         # Vector memory + observations (4-tier RAG)
│       │   ├── tools/          # DOCX/XLSX/PPTX/PDF/OnlyOffice generation
│       │   ├── documents/      # Document studio management
│       │   ├── attachments/    # File upload + OCR processing
│       │   ├── compliance/     # EU AI Act (consent, feedback, bias monitor, audit)
│       │   ├── orchestrator/   # Terminal slot management (Auto-Claude)
│       │   ├── parlant/        # Parlant AI agent proxy
│       │   ├── ingestion/      # URL/text/memory import pipeline
│       │   ├── forms/          # Conversational forms engine
│       │   ├── scheduler/      # Job scheduling (WhiteRabbit)
│       │   ├── marketplace/    # Plugin/skill catalog routes (via marketplace service)
│       │   └── activity/       # Activity logging
│       ├── services/           # Business logic (30+ services)
│       │   ├── ModelRouter.ts              # Rule-based model selection
│       │   ├── SemanticRouter.ts           # Embedding task classification
│       │   ├── ResponseQualityChecker.ts   # Cascade quality assessment
│       │   ├── CircuitBreakerService.ts    # Provider health tracking
│       │   ├── HyDEService.ts              # Hypothetical Document Embeddings
│       │   ├── LLMSyncWorker.ts            # Background LLM sync
│       │   ├── MCPClientManager.ts         # MCP protocol manager
│       │   ├── ParlantProvider.ts          # Parlant integration
│       │   ├── VisionService.ts            # Ollama Vision OCR
│       │   ├── WebSearchService.ts         # Web search integration
│       │   └── document-processing/        # PDF edit session manager
│       └── database/                       # Connection pool + auto-migrations
├── frontend/                   # React 18 + Vite + Tailwind CSS
│   └── src/
│       ├── pages/              # Route pages
│       │   ├── ChatPage.tsx            # Main chat interface
│       │   ├── DocumentsPage.tsx       # Document studio
│       │   ├── ProjectsPage.tsx        # Kanban project management
│       │   ├── AutoClaudePage.tsx      # Autonomous agent dashboard
│       │   ├── MarketplacePage.tsx     # Plugin/skill marketplace
│       │   ├── ParlantPage.tsx         # Parlant agent management
│       │   ├── AITransparencyPage.tsx  # EU AI Act transparency
│       │   ├── SettingsPage.tsx        # User settings + voice config
│       │   └── admin/                  # 20+ admin pages
│       ├── components/         # Reusable UI components
│       ├── hooks/              # Zustand stores (auth, agents, parlant, documents)
│       └── services/           # API client (axios + SSE with routing events)
├── marketplace/                # Standalone marketplace service (Fastify + SQLite)
│   └── src/
│       ├── catalog/            # Catalog service + routes
│       ├── install/            # Install service
│       ├── approval/           # Approval workflow
│       ├── backend/            # Backend client for automated installs
│       ├── kb/                 # Knowledge base integration
│       ├── qdrant/             # Vector search
│       └── database/           # SQLite connection + migrations
├── vscode-extension/           # VS Code companion extension
│   ├── src/                    # Extension entry + 19 commands
│   └── webview-ui/             # React webview bundles (webpack + esbuild)
├── doc-processor/              # Document processing microservice
├── k8s/                        # Kubernetes manifests (Kustomize)
│   ├── backend/                # Deployment + Service + HPA
│   ├── frontend/               # Deployment + Service (Nginx)
│   ├── mariadb/                # StatefulSet + init ConfigMap + schema
│   ├── redis/                  # StatefulSet
│   ├── litellm/                # LiteLLM proxy deployment + configmap
│   ├── parlant/                # Parlant AI service
│   ├── qdrant/                 # Qdrant vector DB
│   ├── marketplace/            # Marketplace service deployment
│   ├── onlyoffice/             # OnlyOffice document server
│   ├── open-webui/             # Open-WebUI deployment
│   ├── tls/                    # TLS certificates
│   ├── storage/                # PersistentVolumes
│   └── kustomization.yaml
├── BUILD.sh                    # Full build pipeline
├── DEPLOY.sh                   # Quick deploy script
└── ROADMAP.md                  # Development roadmap

Configuration

Backend Environment (.env)

# Server
PORT=3000
NODE_ENV=production

# Database (MariaDB)
DB_HOST=mariadb
DB_PORT=3306
DB_USER=enterprise_ai_chat
DB_PASSWORD=your_password
DB_NAME=enterprise_ai_chat
DB_CONNECTION_LIMIT=25

# Redis
REDIS_HOST=redis
REDIS_PORT=6379
REDIS_PASSWORD=your_redis_password

# JWT
JWT_SECRET=your_jwt_secret_min_32_chars
JWT_ACCESS_EXPIRES_IN=15m
JWT_REFRESH_SECRET=your_refresh_secret

# MFA
MFA_BYPASS_EMAILS=                      # comma-separated emails exempt from MFA
TRUSTED_IPS=                            # comma-separated trusted IPs (bypass rate limit)

# AI Providers (or configure via Admin Panel)
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=...

# Ollama (Docker host, NOT Kubernetes)
OLLAMA_BASE_URL=http://10.0.1.1:8086/ollama
OLLAMA_AUTH_KEY=your_ollama_auth_key

# Storage
STORAGE_ROOT=/data/projects
EXTENSION_DIR=/data/projects/extensions

# Encryption (for stored secrets)
ENCRYPTION_KEY=your_32_char_hex_key

# CORS
CORS_ORIGIN=https://your-domain.com

Kubernetes Infrastructure

Deployed on MicroK8s in namespace enterprise-ai-chat:

Service Type Notes
backend Deployment (2 replicas) Fastify API, port 3000
frontend Deployment (2 replicas) Nginx, port 80
mariadb StatefulSet Port 3306, PVC 20Gi
redis StatefulSet Port 6379, PVC 5Gi
qdrant Deployment Vector DB, port 6333
litellm Deployment LLM proxy, port 4000
parlant Deployment AI agent framework, port 8800
marketplace Deployment Plugin catalog, port 3001
onlyoffice Deployment Document server, port 80
open-webui Deployment Alternative UI
doc-processor Deployment Document processing microservice

All external traffic goes through Cloudflare Tunnel. Ingress uses TLS with cert-manager.

Development Commands

# Backend
cd backend
npm run dev          # Dev server with tsx watch
npm run build        # TypeScript compilation → dist/
npm run lint         # ESLint
npm run test         # Vitest unit tests
npx vitest run path/to/test.ts  # Single test file

# Frontend
cd frontend
npm run dev          # Vite dev server (port 5173)
npm run build        # tsc + vite build → dist/
npm run lint         # ESLint
npm run test:e2e     # Playwright E2E tests
npm run test:e2e:prod  # E2E against production

# Marketplace
cd marketplace
npm run dev          # Dev server
npm run build        # TypeScript compilation

# VS Code Extension
cd vscode-extension
npm run build:all    # Build webview + compile extension
npm run webpack      # Production webpack build
npm run webpack:dev  # Dev build with watch
npm run package      # Create .vsix with vsce
npm run release      # Bump version + build + package

# Kubernetes
sudo microk8s kubectl get pods -n enterprise-ai-chat
sudo microk8s kubectl logs -l app=backend -n enterprise-ai-chat
sudo microk8s kubectl rollout status deployment/backend -n enterprise-ai-chat

Security

  • Authentication: JWT (access 15m + refresh 7d) + MFA TOTP + Google OAuth
  • Authorization: Role-based (admin/user) with per-group permissions
  • Input validation: Zod schemas on all API endpoints
  • Rate limiting: Per-IP + per-user limits via @fastify/rate-limit
  • Headers: Helmet CSP, HSTS, X-Frame-Options
  • Network: K8s NetworkPolicy restricts pod-to-pod traffic
  • Secrets: Kubernetes secrets for API keys, never in source
  • Cloudflare: All traffic via Cloudflare Tunnel, IP restrictions via CF-Connecting-IP header

EU AI Act Compliance

  • Art. 50.1 — Mandatory AI disclosure banner on all AI-generated content
  • Art. 50.2 — Synthetic media watermarking (images)
  • Consent management — Granular consent collection and audit trail
  • Bias monitoring — Automated bias detection on model responses
  • Audit logging — Immutable log of all AI interactions
  • AI Transparency page — User-facing explanation of AI system behavior
  • DPIA — Data Protection Impact Assessment documented in DPIA.md

Branches

  • main — Production branch, tracks live deployment.
  • abandoned/* — Archived feature branches preserved for historical reference (no longer maintained):
    • abandoned/archive-legacy
    • abandoned/feat-ai-act-compliance
    • abandoned/feature-agent-framework-v1.6
    • abandoned/feature-document-studio
    • abandoned/feature-v2.0.0-image-gen-voice
    • abandoned/feature-vision-document-pipeline
    • abandoned/feature-vllm-async-document-queue (merged into main)
    • abandoned/feature-vllm-integration
    • abandoned/pre-vllm-migration-backup
    • abandoned/worktree-marketplace

License

Proprietary. See DPIA.md and PRIVACY_POLICY.md for compliance documentation.

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages