AI Customer Support Platform

A scalable, document-aware AI customer support system built to solve real SaaS support problems — not a generic chatbot.

The system answers strictly from company PDFs, classifies user intent, evaluates risk, and escalates to available human agents only when required.

Key Features

• Document-First Answers: All responses come from uploaded client PDFs only
• Per-Client Isolation: Each client has their own vector store and document index
• LangGraph Orchestration: Strict state machine controlling the flow
• Smart Escalation: High-risk intents with high confidence route to humans
• Local Embeddings: Uses all-MiniLM-L6-v2 (no OpenAI dependency)
• Anti-Hallucination: Returns "This information is not available in the provided documentation." when no relevant docs found

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Configure Environment

Create a .env file:

# LLM Providers (at least one required)
GEMINI_API_KEY=your_gemini_key
OPENROUTER_API_KEY=your_openrouter_key

3. Add Documents

Place PDF files in the client directory:

data/documents/{client_id}/
├── product_guide.pdf
├── faq.pdf
└── policies.pdf

4. Run the Server

uvicorn app.main:app --reload

5. Ingest Documents

curl -X POST http://localhost:8000/ingest/demo_client

6. Chat

curl -X POST http://localhost:8000/chat \
  -H "Content-Type: application/json" \
  -d '{"client_id": "demo_client", "user_id": "user123", "message": "How do I reset my password?"}'

API Endpoints

Method	Endpoint	Description
POST	/chat	Process support message
POST	/ingest/{client_id}	Ingest client documents
GET	/documents/{client_id}/status	Get document store status
POST	/agents/register	Register human agent
PUT	/agents/{agent_id}/status	Update agent status
GET	/health	Health check
GET	/metrics	Platform metrics

LangGraph Flow

intent_classification → cache_check → [cached?]
                                         ↓ no
                              document_retrieval
                                         ↓
                              relevance_validation → [clarify?]
                                         ↓ no
                              confidence_scoring
                                         ↓
                              escalation_decision → [escalate?]
                                    ↓ no              ↓ yes
                              answer_generation   escalation_response
                                         ↓
                                  cache_response → END

🧪 Example Response

{
  "reply": "Account deletion requests must be submitted to support as per the Terms.",
  "intent": "account_deletion",
  "confidence": 0.82,
  "escalated": false
}