Technical architecture
A straightforward, production-shaped path from request to response — a Next.js frontend, a FastAPI backend, and a LangGraph agent that decides when to retrieve knowledge, call a tool, or just answer.
Next.js 16
App Router · streaming UI over SSE
FastAPI
REST + Server-Sent Events
LangGraph Agent
Tool-calling loop, per-session memory
LLM
OpenAI gpt-5.4-nano
RAG
FAISS + text-embedding-3-small
Tools
Booking, pricing, hours, leads
PostgreSQL
Leads, chats, analytics, documents
Tech stack
Frontend
- Next.js 16 (App Router)
- TypeScript
- Tailwind CSS v4
- shadcn/ui
- Framer Motion
Backend
- FastAPI
- Python 3.13
- SQLAlchemy
- Pydantic
AI / Agent
- LangGraph
- LangChain
- OpenAI gpt-5.4-nano
- text-embedding-3-small
- FAISS
Infrastructure
- PostgreSQL
- Docker Compose
Engineering highlights
- Token-level streaming over SSE, parsed manually on the client (no vendor lock-in to a specific chat SDK)
- Tool-calling agent that decides between RAG search, structured business tools, or a direct reply
- Hallucination guard: a similarity threshold plus an explicit system-prompt rule to refuse ungrounded answers
- Per-session conversation memory via a LangGraph checkpointer, keyed by session id
- Conversational multi-turn slot-filling for lead capture (no rigid form, no repeated questions)
- PDF upload → chunk → embed → live FAISS index update, searchable without a server restart
- Admin analytics computed live from logged queries — no precomputed/mocked metrics