ABC Customer Support Automation

Live Demo: https://adroite2.onrender.com

This project routes customer questions through a LangGraph workflow for Sales, Technical Support, Billing, and Account Support. It retrieves company guidance, keeps customer history in SQLite, pauses risky requests for a supervisor, and checks every response before returning it to the customer.

Workflow

The graph follows this order:

1. classify_intent detects the department and checks for high-risk language.
2. Memory questions go directly to recall_memory.
3. Other questions go through retrieve_context and then to one department agent.
4. Refunds, cancellations, account closures, compensation, and management escalations pause at human_review.
5. The supervisor checks the draft before finalize stores the interaction and returns the response.

LangGraph checkpoints and the readable support_interactions table share the same SQLite database. A customer ID is also used as the LangGraph thread ID, so the workflow keeps each customer's state separate.

State

SupportState contains:

• request and customer IDs
• current query and classified intent
• high-risk flag and approval status
• retrieved context and document sources
• department draft, reviewer feedback, and final response
• accumulated LangChain messages

RAG

The local retriever loads the four Markdown files in knowledge-base. It scores each document with term frequency and inverse document frequency, then supplies the three best documents to the selected support agent. The agent is instructed to answer only from that retrieved context.

This keeps the sample project small and avoids a separate vector database. New company guidance can be added as another Markdown file.

Human review

The graph calls LangGraph interrupt() for every high-risk request. The graph and its draft are saved in SQLite while execution is paused. The Streamlit page shows Approve and Reject controls. Either decision resumes the same thread with Command(resume=...).

An approval means the request can move to the next processing step. It does not claim that money was transferred or an account was changed. A rejection returns the supervisor's reason without implying approval.

Setup

Python 3.10 or later is required.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
export GROQ_API_KEY="your-key"

The Groq key is read only from the environment. Do not add it to .env.example or commit a local .env file.

Start the web interface:

streamlit run app.py

Run the five required queries in the terminal:

python support_system.py --demo --db demo-memory.db

Run one query:

python support_system.py \
  --customer customer-001 \
  --query "I forgot my account password."

Without GROQ_API_KEY, the application uses fixed local responses. Routing, retrieval, SQLite memory, and human review still run through the same LangGraph. This mode is useful for checking the workflow without spending API quota.

Features & Agent Routing

Query	Route	Human approval
What are the pricing plans available for your software?	Sales	No
I forgot my account password.	Account	No
My application crashes whenever I upload a file.	Technical Support	No
I need a refund for my annual subscription.	Billing	Yes
What was my previous support issue?	Memory Recall	No

The fifth query must use the same customer ID as an earlier query. The included demo does this automatically.

SQLite schema

schema.sql creates the readable interaction table. LangGraph creates its own checkpoint tables in the same memory.db file on first run.

sqlite3 memory.db < schema.sql

Runtime database files are ignored by Git.

Deployment

render.yaml defines a Render web service. Create a Blueprint from the repository and set GROQ_API_KEY in the Render dashboard. The free filesystem is temporary, so a production deployment should mount a persistent disk for memory.db.