ResolveX (LangGraph + Pinecone + Postgres)

Live: https://customer-support-ewuu.onrender.com/static/admin-dashboard.html

This project is an AI-powered customer support resolution system with:

• Email ingestion from company inbox (IMAP)
• LangGraph agent to decide whether an email can be auto-resolved
• Pinecone retrieval from company policies/documents
• Automatic reply to customers for resolvable queries (SMTP)
• Human escalation path for unresolved queries
• Multi-company onboarding with company admin login
• Role-based auth (company_admin, human_agent)
• Postgres storage for unresolved tickets
• Support team web console for login and manual ticket resolution

Architecture

1. Poll unread emails from inbox.
2. For each email, run LangGraph workflow:
   • Retrieve company-specific policy context from Pinecone namespace.
   • Ask LLM if issue can be resolved using policy context.
   • If resolvable, send AI reply to customer.
   • Else, create unresolved ticket in Postgres for that company.
3. Human support agents login and resolve escalated tickets.

Stack

• Backend: FastAPI
• Agent orchestration: LangGraph
• LLM: Gemini via langchain-google-genai
• Vector DB: Pinecone
• Relational DB: PostgreSQL
• UI: Static HTML/CSS/JS served by FastAPI

Project Structure

• backend/app/main.py: FastAPI entrypoint
• backend/app/agent/graph.py: LangGraph workflow
• backend/app/pinecone_client.py: Pinecone retrieval helper
• backend/app/email_client.py: IMAP/SMTP integration
• backend/app/services/email_processor.py: end-to-end email processing
• backend/app/routers/*: APIs for auth, ingestion, and tickets
• backend/app/static/*: support team dashboard
• backend/scripts/index_policies.py: policy indexing script

Setup

1) Start local PostgreSQL

Ensure PostgreSQL is installed and running on your machine, and create database customer_care.

Example with psql:

psql -U postgres -c "CREATE DATABASE customer_care;"

2) Python environment and dependencies

cd backend
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

3) Configure environment

copy .env.example .env

Update .env values:

• GEMINI_API_KEY
• GEMINI_MODEL
• GEMINI_EMBEDDING_MODEL
• PINECONE_API_KEY
• PINECONE_INDEX_NAME
• DATABASE_URL
• EMAIL_USER, EMAIL_PASSWORD
• IMAP_HOST, SMTP_HOST if different from defaults
• HUMAN_SUPPORT_EMAIL

4) Index company docs/policies to Pinecone

Place docs under backend/data (.txt or .md) and run:

python scripts/index_policies.py --docs ./data

5) Run API

uvicorn app.main:app --reload

Open: http://localhost:8000

Customer Care Login

1. Register human agent via POST /api/auth/register/human-agent using username + company_email + password.
2. Human login via POST /api/auth/login/human using username + password.
3. Customer care agents see and resolve unresolved tickets only for their own company.

The built-in web console now supports:

• simple customer care login with username and password
• unresolved ticket listing and resolution

API Endpoints

• POST /api/auth/register/human-agent - human agent registration (username + company email + password)
• POST /api/auth/login/human - human agent login (username + password)
• GET /api/auth/me - get current authenticated user profile
• POST /api/ingest/poll - poll current company inbox + process unread emails (auth required)
• GET /api/tickets - list company unresolved/resolved tickets (auth required)
• PATCH /api/tickets/{ticket_id} - resolve/update ticket (auth required)
• GET /health - health check

Reply Ownership Rule

• Auto-resolved messages are replied by AI and do not create unresolved tickets.
• Escalated tickets are replied by human agents on resolve.
• A ticket prevents duplicate customer replies from multiple responders.

Notes for Production

• Add scheduler/worker (e.g., APScheduler, Celery, or cron) to invoke /api/ingest/poll automatically.
• Replace default login with proper user management + password reset.
• Add audit logs and role-based permissions.
• Secure email credentials via secret manager.
• Add robust JSON parsing in graph.py and guardrails for AI outputs.

Render Deployment (Production)

Use AI_Services/app as the service root directory.

• Root directory: AI_Services/app
• Build command: pip install -r requirements.txt
• Start command: uvicorn main:app --host 0.0.0.0 --port $PORT
• Health check path: /health

Required environment variables:

• DATABASE_URL (from Render PostgreSQL connection string)
• SECRET_KEY
• GEMINI_API_KEY
• PINECONE_API_KEY
• PINECONE_INDEX_NAME

Recommended production values:

• DEBUG_MODE=false
• AUTO_POLL_ENABLED=false