ENTERPRISE AI DOCUMENT CHATBOT SYSTEM -- RAG ARCHITECTURE (Corporate Scale) ───────────────────────────────────────── PROJECT OVERVIEW ───────────────────────────────────────── We are building a production-grade AI chatbot system powered by Retrieval-Augmented Generation (RAG) for a organisation in Malaysia. The system must handle thousands of documents (PDFs, Word, Excel), live database queries, and serve two distinct user types -- public citizens and internal staff. This is NOT a simple chatbot. This is an enterprise knowledge management system with strict data security, role-based access, bilingual support (Bahasa Malaysia + English), and full audit compliance. ───────────────────────────────────────── CORE REQUIREMENTS ───────────────────────────────────────── 1. DOCUMENT INGESTION PIPELINE - Support PDF, Word (.docx), Excel (.xlsx/.csv), and plain text files - Handle 10,000+ documents efficiently with batch processing - Automatic PII detection and redaction before embedding (Malaysian IC, phone numbers, emails, addresses) - Smart chunking strategy (sentence-aware, 150-250 words per chunk with overlap) - Document tagging system: public / internal / confidential - Incremental ingestion -- only re-process changed or new documents - Support for scanned PDFs via OCR (Tesseract or equivalent) - Progress tracking and ingestion logs 2. VECTOR DATABASE AND RETRIEVAL - Qdrant as primary vector database (self-hosted) - Support for 1M+ vectors with fast retrieval (under 200ms) - Hybrid search: semantic (vector) + keyword (BM25) combined scoring - Metadata filtering by document tag, department, date range - Re-ranking of retrieved chunks for accuracy - Separate collections per document category - Database integration: connect to PostgreSQL/MySQL and answer queries from live data 3. CITIZEN BOT (Public Facing) - Clean, professional portal UI - Bilingual: Bahasa Malaysia + English (auto-detect or manual toggle) - Answers STRICTLY from public-tagged documents only - Hardcoded access filter -- cannot be overridden by user input - Prompt injection protection - Rate limiting (10 requests per minute per IP) - Graceful fallback message when answer not found - Mobile responsive - No authentication required 4. STAFF BOT (Internal) - Secure login with JWT authentication - Role-based access control: Clerk / Manager / Director / Admin - Each role sees documents matching their clearance level - Full conversation memory per session - Report generation: user types natural language, system generates SQL, returns formatted data - Export reports as PDF and Excel - Audit log: every query logged immutably (user, timestamp, question, sources, response) - Staff can view own logs; Directors can view all logs 5. LLM INTEGRATION - Primary: OpenAI GPT-4o-mini via API/Any suggested by you - Fallback option: local Ollama (Mistral/Llama) for air-gapped environments - Temperature 0.0 for factual accuracy - Strict document-only answering -- no hallucination from general knowledge - Response must always cite source document and page number - Support for streaming responses 6. SECURITY - All data stays on-premise (no third-party data storage) - PII never stored in vector database - SQL injection protection on report generation - Read-only database credentials for bot queries - HTTPS enforcement in production - Input sanitisation on all endpoints - OWASP-compliant API design 7. INFRASTRUCTURE - Docker + Docker Compose for all services - Services: vector DB, citizen API, staff API, frontends, PostgreSQL, Redis cache - Environment-based configuration (.env) - Health check endpoints - Horizontal scaling support - Logging and monitoring hooks (compatible with Grafana/Prometheus) 8. TECH STACK (preferred, open to discussion) - Backend: Python FastAPI or PHP (Laravel) - Frontend: React.js - Vector DB: Qdrant - Relational DB: PostgreSQL - Cache: Redis - LLM: OpenAI API + Ollama fallback - Embeddings: text-embedding-3-small (OpenAI) or nomic-embed-text (Ollama) - Containerisation: Docker - Auth: JWT ───────────────────────────────────────── DELIVERABLES ───────────────────────────────────────── - Complete source code (GitHub repo, clean commits) - Citizen Bot -- fully functional frontend + API - Staff Bot -- fully functional frontend + API with all role levels - Ingestion pipeline script (handles 10,000+ docs) - Docker Compose setup (one command to run everything) - Database schema + migration scripts - Comprehensive README with Windows + Linux setup instructions - API documentation (Swagger/OpenAPI) - Security audit report - Performance benchmark (query response time with 100k+ vectors) - 2 weeks post-delivery bug fix support ───────────────────────────────────────── IDEAL CANDIDATE ───────────────────────────────────────── - Strong experience with RAG architecture and vector databases - Built production AI systems, not just demos or tutorials - Familiar with Malaysian government context or similar multilingual government deployments - Security-conscious -- understands data privacy requirements - Can work independently and communicate clearly in English - Portfolio of similar AI/chatbot projects required - Experience with large-scale document processing (10k+ files) is a strong plus ───────────────────────────────────────── BUDGET AND TIMELINE ───────────────────────────────────────── Budget: Open to proposals -- please provide a detailed breakdown Timeline: ASAP Engagement: Fixed price preferred, milestone-based payments Milestones suggested: - Milestone 1 (20%): Ingestion pipeline + Qdrant setup working with 1,000 test docs - Milestone 2 (50%): Both chatbot APIs functional with correct document-only answering - Milestone 3 (20%): Both frontends complete, report generation, audit logging - Milestone 4 (10%): Docker deployment, documentation, final testing ───────────────────────────────────────── TO APPLY ───────────────────────────────────────── Please include in your proposal: 1. A brief description of a similar RAG or document chatbot system you have built 2. Your recommended tech stack and any changes to the above 3. Estimated timeline and milestone breakdown 4. How you would handle 10,000+ document ingestion efficiently 5. Your approach to preventing the bot from answering outside the documents Proposals without relevant experience will not be considered. This is a serious, funded project for a client. We are looking for a professional who delivers production-quality work, not a proof of concept. We are also willing to pay for a proof of concept first. Upon successful, we will pay for a full development.