Overview | Chatbot in Action | Quick Start | Key Findings | Architecture | Reports | Bug Fix
A domain-specific RAG chatbot built on Azure AI Foundry, customised for Melbourne logistics operations using Linfox Australia as the domain. Allows operations staff to query depot locations, delivery zones, freight types, shift times, delay reasons and escalation procedures through natural language.
Forked from Azure-Samples/get-started-with-ai-chat and extended with custom data ingestion, infrastructure fixes and systematic behaviour analysis.
Built to investigate emergent RAG behaviour in a production-like environment — how chunking boundaries, prompt versions, and retrieval configurations interact to produce unexpected results in real conditions. Linfox Melbourne logistics operations was chosen as the domain to make the failure modes operationally meaningful rather than theoretical.
Five queries demonstrating correct retrieval, chunking failure, conversation history behaviour, information injection and out of scope handling. Full analysis in the RAG Behaviour Analysis Report.
Prerequisites: Python 3.10+, Node.js 18+, Azure subscription, Azure Developer CLI (azd)
# 1. Clone the repo
git clone https://github.com/deesk/linfox-rag-chatbot-azure.git
cd linfox-rag-chatbot-azure
# 2. Install dependencies
pip install -r src/api/requirements.txt
# 3. Deploy to Azure
azd up
# 4. Shut down when done
azd down --purgeDeployment takes approximately 25 minutes.
Note: Pre-generated embeddings are included in the repo. Only run
python scripts/build_embeddings.pyif you modify the knowledge base data.
The Microsoft template has a built-in embedding process for its own sample data. To support custom domain data, a reusable pipeline was built: scripts/build_embeddings.py
This script reads .txt files from src/static/data/, generates embeddings using text-embedding-3-small, saves embeddings.csv to src/api/data/, and deletes the existing Azure Search index so it is recreated fresh on next deploy.
To use your own domain data, replace src/static/data/melbourne-logistics.txt then run:
python scripts/build_embeddings.pyThen redeploy:
azd up
A standard RAG pipeline — user query is embedded, matched against the logistics knowledge base in Azure AI Search, top 5 chunks injected into GPT-4o-mini alongside conversation history to generate a response.
Full architecture and data pipeline details in the RAG Behaviour Analysis Report.
RAG quality depends primarily on data structure, chunking strategy and prompt engineering, not the AI model itself.
What worked:
- Depot locations, shift times and escalation procedures returned correctly when content landed within a single retrieval chunk
- Out of scope queries handled gracefully without explicit instruction
- Conversation history supplemented incomplete RAG retrieval in real user sessions
What failed:
- System returned 3 of 4 freight types with full confidence. Dangerous goods missing due to a chunk boundary. No error signal to the user
- Delivery and loading zones contaminated each other because they shared Melbourne suburb names in content
- User injected false delay reasons. System accepted them, used real knowledge base data to make them sound credible, and carried them forward even when challenged
Key insight: The most dangerous failure is not a sorry response. It is a confident wrong answer. A system that sounds certain while missing critical information gives the user no signal to question it.
Full investigation: RAG Behaviour Analysis Report
| Component | Technology |
|---|---|
| AI Model | GPT-4o-mini (Azure AI Foundry) |
| Embeddings | text-embedding-3-small (100 dimensions) |
| Vector Search | Azure AI Search |
| Backend | Python, Gunicorn, Azure Container Apps |
| Frontend | React (Microsoft template) |
| Infrastructure | Azure Bicep, Azure Developer CLI (azd) |
| Data pipeline | Custom Python (scripts/build_embeddings.py) |
RAG Behaviour Analysis Report Systematic investigation across 6 parts and 8 findings covering chunking failures, vector contamination, conversation history behaviour, prompt injection and system prompt security. Each finding documented with evidence, methodology and root cause analysis.
Cost Analysis Real Azure spend breakdown across 16 active development days (AU$2.59 total). Covers provisioned vs consumption billing, idle cost observations and production considerations.
A blocking bug was identified and fixed in this fork from the upstream Microsoft template.
Internal Server Error on first load (Issue #129)
After a successful azd up deployment, the app returned an Internal Server Error on every page load immediately. The root cause was a TemplateResponse syntax in src/api/routes.py incompatible with the current Starlette version on Python 3.13.
Fixed by updating the argument format:
# Before (broken)
return templates.TemplateResponse(
"index.html",
{"request": request}
)
# After (fixed)
return templates.TemplateResponse(
request=request,
name="index.html"
)Portfolio proof of concept. Not affiliated with or endorsed by Linfox Australia.