Gartech Vercel Bot

The core architecture follows a domain‑driven modular monolith pattern. The /app directory hosts the Next.js UI and API routes, using route groups like (chat) and (auth) to isolate concerns without affecting URLs. The primary chat endpoint POST /api/v1/chat (also exposed as /api/chat for streaming) receives a user message, invokes queryQdrant to fetch relevant knowledge base vectors, augments the system prompt, and calls the Gemini model, streaming the response back to the client. Behind the scenes, the extractAndSaveLeadInfo routine parses the conversation for lead attributes (name, phone, capacity, land size) and persists them via Mongoose schemas.

Data ingestion is handled by a PDF parsing pipeline that chunks documents, stores raw text in FAQ/KnowledgeItem collections, and pushes embeddings to Qdrant. The vector store is queried in real‑time to provide RAG capabilities, ensuring answers are grounded in the uploaded knowledge base. Security is scaffolded with NextAuth.js (MongoDB adapter) for session management, JWT‑based strategies, and planned middleware for authentication, rate‑limiting (Redis‑backed), and centralized error handling.

The backend resides under /backend/src with clear layers: config for env validation, routes for API gateways, controllers for request handling, and services for business logic (AI, conversation state, WhatsApp webhook processing, CRM integration). Persistence is abstracted via lib/db/connection.ts and lib/db/queries.ts, providing typed access to Chat, Message, Lead, and knowledge collections. Telemetry and analytics hooks are earmarked for future integration, feeding metrics into an admin dashboard at /admin.

Deployment targets Docker containers on Vercel (frontend) and Google Cloud Run (backend), enabling horizontal scaling. The codebase includes TypeScript typings throughout, ensuring type safety across API contracts and database models. Continuous development is supported by utility scripts such as scripts/make-admin.js for role management.