Hai Phong Radio and Television Station (BPTTH)
An AI-powered chat assistant that helps the Advertising & Business Department look up pricing, calculate advertising costs, and get instant answers from internal documents — without digging through PDFs manually.
The Advertising & Business Department at BPTTH works with multiple overlapping price tables (THP, HP channels, radio, digital, documentaries), each governed by separate official decisions (QĐ 413, 414, 415). Staff regularly need to:
- Find the right unit price for a given time slot and customer type
- Calculate total contract costs including discounts
- Determine which pricing decision applies (Hai Phong client vs. out-of-province)
Doing this manually is slow and error-prone, especially under client pressure.
A conversational AI assistant backed by a Knowledge Graph built from the station's actual documents. Instead of keyword search, the system understands intent and retrieves contextually relevant information.
Three modes:
- Q&A mode — answers pricing and policy questions with source citations
- Calculate mode — computes exact costs using dedicated Python pricing tools (LLM handles presentation, not arithmetic)
- Quote mode — generates professional advertising quote proposals in Word format with pricing details and customer information
Visual Architecture:
The core of the system. Source documents are processed into a Knowledge Graph where:
- Text is split into chunks and embedded with a Vietnamese embedding model
- Entities (time slots, prices, channels, customer types) and their relationships are extracted by an LLM
- Three indexes enable hybrid search: vector, fulltext, and entity-level
Built with a customized version of Neo4j LLM Graph Builder, I forked this repo then refactor it with some strategies and config. Adapted for Vietnamese and BPTTH-specific content, you can contact me via email pminhtu24@gmail.com if u interested in.
Quote mode extracts customer information and pricing details from the conversation context to automatically generate professional advertising quote proposals in Word format (DOCX):
- Extracts customer name, contact, requested time slots, and channels
- Retrieves applicable pricing rules from Neo4j Knowledge Graph
- Generates formatted pricing table with pricing decision references (QĐ 413, 414, 415)
- Creates downloadable Word document ready for client delivery
- Files stored in
backend/generated_quotes/with session-based naming for easy cleanup
Example: "Hãy tạo báo giá cho khách hàng ABC với giá gói HP1 trong tháng..." → Generated Word doc with full pricing breakdown
| Layer | Technology |
|---|---|
| Backend | Python, FastAPI, LangGraph |
| LLM | GPT-OSS-120B |
| Embedding | AITeamVN/Vietnamese_Embedding_v2 (self-hosted via Infinity) |
| Knowledge Graph | Neo4j (AuraDB) |
| Frontend | React, Vite, TypeScript, Tailwind CSS |
broadcast-AI-application/
├── backend/
│ ├── main.py # FastAPI entry point
│ └── app/
│ ├── api/routes/
│ │ ├── chat.py # POST /api/chat, /stream, session routes
│ │ └── health.py # GET /health
│ ├── config/
│ │ ├── settings.py # Pydantic settings — reads backend/.env
│ │ └── constants.py # Cypher retrieval queries, LLM prompts, search params
│ ├── graph/ # LangGraph workflow
│ │ ├── state.py # ChatState definition
│ │ ├── nodes.py # Node functions (load_session, classify, retrieve, generate, ...)
│ │ ├── edges.py # Conditional routing logic
│ │ └── workflow.py # Graph compilation + invoke_graph()
│ ├── schemas/
│ │ └── chat.py # Pydantic request / response models
│ └── services/
│ ├── retriever.py # GraphRAG Hybrid search (vector + graph + fulltext)
│ ├── llm.py # LLM singleton (Viettel AI)
│ ├── session.py # Neo4j session CRUD
│ ├── pricing_tools.py # Python pricing calculators (pure functions)
│ └── tools.py # LangChain StructuredTool wrappers
│
├── frontend/
│ └── src/
│ ├── components/
│ │ ├── ChatInput.tsx # Input box + send button
│ │ ├── Message.tsx # Message bubble with citations + KaTeX math
│ │ └── Sidebar.tsx # Session list panel
│ ├── pages/
│ │ └── ChatPage.tsx # Main chat page
│ ├── hooks/ # Custom React hooks
│ ├── types/ # TypeScript type definitions
│ └── lib/ # Shared utilities
│
├── benchmark/
│ └── run.py # RAGAS evaluation benchmark
├── models/
└── local_model_AITeamVN_Vietnamese_Embedding_v2/ # Local embedding model
Requirements: Python ≥ 3.10, Node.js ≥ 18, pnpm, uv, Docker, Neo4j with APOC
# 1. Install dependencies
uv sync
cd frontend && pnpm install && cd ..
# 2. Configure environment
cp backend/.env.example backend/.env # fill in NEO4J_*, INFINITY_URL, VIETTEL_*
# 3. Start embedding server
docker run -d -p 7997:7997 --gpus all \
-v ./models/local_model_AITeamVN_Vietnamese_Embedding_v2:/models/Vietnamese_Embedding_v2 \
michaelf34/infinity:latest v2 \
--model-id /models/Vietnamese_Embedding_v2 --port 7997
# 4. Build the Knowledge Base
# Upload documents via Neo4j LLM Graph Builder → point to your Neo4j instance
# 5. Run backend
PYTHONPATH=backend uv run uvicorn backend.main:app --reload --port 8000
# 6. Run frontend
cd frontend && pnpm dev # http://localhost:5173| Endpoint | Method | Description |
|---|---|---|
/api/chat |
POST | Chat (full response) — supports qa, calculate, quote modes |
/api/chat/stream |
POST | Chat with SSE token streaming |
/api/chat/{session_id} |
DELETE | Delete a session and its quotes |
/api/chat/sessions |
GET | List all sessions |
/api/chat/{session_id}/history |
GET | Get message history for a session |
/api/quotes/{filename} |
GET | Download generated quote (Word document) |
/api/quotes/{session_id} |
DELETE | Clean up quotes for a session |
/health |
GET | Health check |
Evaluated 4 retrieval strategies on 50 custom test questions using RAGAS metrics.
| Retrieval Mode | Median Latency | Mean Latency | Answer Relevancy | Context Precision | Context Recall | Faithfulness |
|---|---|---|---|---|---|---|
| Fulltext only | 146ms | 149ms | 0.649 | 1.000 | 1.0 | 1.0 |
| Vector only | 2,032ms | 2,695ms | 0.789 | 1.000 | 1.0 | 1.0 |
| Graph + Vector | 1,406ms | 1,961ms | 0.855 | 1.000 | 1.0 | 1.0 |
| Graph + Vector + Fulltext | 1,523ms | 1,839ms | 0.884 | 1.000 | 1.0 | 1.0 |
Key takeaways:
- Hybrid (Graph + Vector + Fulltext) achieves the highest answer relevancy (+11.9% over vector-only) while being ~25% faster at median latency
- Fulltext-only is fastest but has the lowest relevancy — poor at semantic queries
- Context Precision / Recall / Faithfulness are near-perfect across all modes, meaning retrieved chunks are always relevant and answers stay grounded
Design decision: The hybrid mode adds ~1.4s over fulltext-only — a deliberate trade-off for significantly higher answer relevancy. At ~1.5s median retrieval latency, the UX remains responsive and acceptable for a domain-specific internal tool.
Internal use — Hai Phong Radio and Television Station · v0.1.0 · April 2026