Chat Support Agent (RAG)

Production-style customer support backend built with FastAPI, OpenAI, Pinecone, Redis, and Supabase.

The service answers user questions from internal documentation using Retrieval-Augmented Generation (RAG). If confidence is low, it automatically escalates to a support ticket workflow.

The problem it solves

Every support team has the same issue: customers ask the same questions over and over. How do I cancel? What payment methods do you accept? How do I reset my password? A human has to read each one, find the answer in some internal doc, and type a response. It's repetitive, slow, and expensive. This agent handles those automatically. The tricky questions — the ones that actually need a human — get escalated to a ticket with the full conversation context already attached.

Why This Project

This project demonstrates practical backend AI engineering patterns that are relevant for real products:

• RAG pipeline for grounded responses
• Async API design with FastAPI
• Cache-aside strategy with Redis
• Automated fallback from AI response to ticket creation
• Clean API contracts using Pydantic models
• Testable architecture with service isolation and mocks

Architecture

flowchart TD
    A[POST /api/v1/chat\nChatRequest] --> B[chat_endpoint]
    B --> C{cache_service.get\nMD5 query hash}
    C -- Cache Hit --> D[ChatResponse\nfrom_cache: true]
    C -- Cache Miss --> E[RAGGenerator.generate\nquery + history]
    E --> F[DocumentRetriever.retrieve\nget_embedding - OpenAI]
    F --> G[index.query\nPinecone top_k=5]
    G --> H{query_answer\nscore >= threshold}
    H -- can_answer: true --> I[_build_context\ndocs + scores]
    I --> J[chat.completions.create\nGPT-4o-mini + SYSTEM_PROMPT]
    J --> K[cache_service.set\nTTL: 3600s]
    K --> L[ChatResponse\ncan_answer: true]
    H -- can_answer: false --> M[ticket_service.create_ticket\nTicketData]
    M --> N[SupabaseBackend\nPOST support_tickets]
    N --> O[ChatResponse\nticket_id + ticket_created: true]

Loading

Tech Stack

• Backend: FastAPI, Uvicorn, Pydantic
• LLM: OpenAI (chat + embeddings)
• Retrieval: LangChain + Pinecone
• Caching: Redis
• Ticket persistence/integration: Supabase (plus Zendesk/HubSpot-ready service layer)
• Testing: pytest, pytest-asyncio, httpx

Project Structure

app/
  main.py                 # FastAPI app entrypoint
  api/routes.py           # API endpoints
  core/config.py          # Settings and environment config
  rag/embeddings.py       # Embedding + indexing logic
  rag/retriever.py        # Semantic retrieval
  rag/generator.py        # Response generation logic
  services/cache_service.py
  services/ticket_service.py

data/docs/                # Knowledge base documents
scripts/ingest_docs.py    # Document ingestion pipeline
tests/test_api.py         # API tests
requirements.txt

Quick Start

1. Clone and install

git clone https://github.com/DavidFSantillan/Chat-Support-Agent.git
cd Chat-Support-Agent

python -m venv .venv
# Windows PowerShell
.\.venv\Scripts\Activate.ps1
# macOS/Linux
# source .venv/bin/activate

pip install -r requirements.txt

2. Configure environment variables

Create a .env file in the project root:

OPENAI_API_KEY=sk-...
PINECONE_API_KEY=...
SUPABASE_URL=https://<project>.supabase.co
SUPABASE_KEY=...
REDIS_URL=redis://localhost:6379

3. Start Redis locally

docker run -d -p 6379:6379 redis:alpine

4. Ingest documentation

python scripts/ingest_docs.py

5. Run the API

uvicorn app.main:app --reload --port 8000

• Swagger UI: http://localhost:8000/docs
• Health check: http://localhost:8000/api/v1/health

API Example

Request

curl -X POST http://localhost:8000/api/v1/chat \
  -H "Content-Type: application/json" \
  -d '{
    "message": "How can I cancel my subscription?",
    "user_email": "customer@example.com",
    "user_name": "Jane Doe"
  }'

Typical response

{
  "conversation_id": "conv_abc123def456",
  "answer": "To cancel your subscription, go to Settings > Billing...",
  "can_answer": true,
  "sources": ["faq.md"],
  "confidence": 0.92,
  "ticket_created": false,
  "ticket_id": null,
  "from_cache": false
}

Testing

pytest -m tests/ -v

With coverage:

pytest -m tests/ -v --cov=app --cov-report=html