AI Customer Support Agent
A RAG + tool-using support agent that answers policy questions and performs controlled order actions from a local, self-contained deployment.
Built by dhyuti · AI Engineering Associate
The system, up close
Real captures from the running system — the demo, screenshots and architecture come first, then the deep dive.
Demo
Live support agent handling policy and order questions, including a controlled ticket action.
The demo shows grounded policy answers plus controlled order operations.
Why this needed to be built
Support teams field repetitive policy and order questions that drown agents and slow response times, while order lookups and ticket creation still need a human in the loop.
Off-the-shelf bots either hallucinate policy or can't safely take actions; teams need grounded answers plus controlled operational tools.
What we built, and how it changes the workflow
We built a Streamlit support agent that combines document-grounded RAG with controlled tools, running fully locally on Ollama so it stays self-contained and cheap to evaluate.
The agent retrieves answers from a Novacart policy knowledge base (ChromaDB) and can perform controlled operations — order lookup, ticket creation, status checks, human escalation — through predefined tools, while the LLM focuses on language and reasoning (llama3.1:8b chosen over 3b for better policy reasoning).
A local SQLite store holds operational data; unsupported questions (e.g. 'deliver to Antarctica?') are answered honestly as out-of-knowledge rather than invented. Nine validation scenarios all passed.
Core design principle
- Knowledge retrieval and operational actions are separated — the LLM reasons, tools act
- Local Ollama deployment keeps the MVP self-contained and private
- Honest uncertainty — out-of-scope questions are not fabricated
What it can do
A focused set of capabilities that define the system — no marketable filler, only what actually runs.
Policy Q&A (RAG)
Grounded answers retrieved from the Novacart policy knowledge base.
Order lookup
Retrieve order status and details via a controlled tool.
Ticket creation
Open support tickets for damaged / problem orders.
Ticket status
Check the status of an existing ticket.
Human escalation
Route cases the agent shouldn't handle to a human.
Uncertainty handling
States when the knowledge base can't answer instead of guessing.
Local deployment
Runs on Ollama with ChromaDB and SQLite — no cloud dependency.
The end-to-end workflow
The system processing path, in order.
- 1
Ask a question
A customer or agent poses a policy or order question in the chat.
- 2
Retrieve context
Relevant policy passages are pulled from ChromaDB.
- 3
Reason
llama3.1:8b generates a grounded answer or decides to use a tool.
- 4
Act (if needed)
Order lookup, ticket creation / status, or escalation via predefined tools.
- 5
Respond
The agent returns the answer or action result to the chat.
- 6
Validate
Behaviour is checked against nine representative scenarios.
Stack & tech choices
- App
- Streamlit
- Orchestration
- LangChain
- Local LLM
- Ollama, llama3.1:8b
- Retrieval & data
- ChromaDB, SQLite
Tested, verified and honestly scoped
Validation & tests
Validated against nine representative customer-support scenarios — all passed (100% for the defined set).
- Return policy, returns within 10 days, non-returnable items — PASS
- Order NC1002 / NC1001 lookup — PASS
- Damaged-order ticket creation, ticket status — PASS
- Human escalation — PASS
- 'Antarctica delivery' unsupported question handled honestly — PASS
Security & hardening
- Local deployment keeps data on-machine
- Operational actions gated behind predefined tools, not free-form
- Knowledge separated from actions to limit hallucinated operations
Honest limitations
AI Customer Support Agent is an internally developed engineering prototype. We state limitations rather than overclaim production readiness.
- Operates only on the configured knowledge base and local data
- MVP for local demo, not production-scale
- No admin KB management UI, auth, or cloud observability yet
Deep-dive references
Documentation
Technical report, architecture, validation set and policy source.
Where this can go next
Realistic next steps we've scoped — not aspirational marketing copy.
- Admin workflow to update / re-index knowledge
- Authentication and role-based access control
- Larger, versioned customer-support test set
- Monitoring, tracing and production observability
- Production database and deployment
- More controlled tools with stronger permissions
Have a similar workflow?
Zenera Labs can build a custom version for your business.
This is one of dhyuti's internally developed systems. Tell us about your requirement and we'll scope a tailored build around it.
Other systems we've built internally
AI Contract Analysis & Risk Detection
Faster document review with traceable evidence — upload a contract and get risk findings with source-verified citations.
naman · Engineering Portfolio
View case study
SentinelAI — Security Log Anomaly Detection
Risk-scored alerts from security events — a full-stack Mini-SIEM with ML detection, attack-chain correlation and explainable alerts.
pragna · Engineering Portfolio
View case study
AI Data Analyst & Business Intelligence Agent
Evidence-grounded business intelligence from spreadsheets — upload CSV/XLSX, get analysis, visualizations and explained answers.
priyanka · Engineering Portfolio
View case study