Ecommerce AI chatbot on Azure OpenAI

Overview

Saba wanted visitors to get quick answers about products, orders, the company and policies. The assistant searches an index of approved question-and-answer entries and crawled site pages, returns a stored answer word for word when it is clearly right, and otherwise drafts a reply from the retrieved passages or says it cannot answer. It replaced an earlier bot whose answers were shipped to every visitor's browser. We built the chat engine inside the storefront, the nightly ingestion job, the chat widget and the test suite.

The challenge

Supplement wording carries regulatory risk, and similarity scores alone cannot tell a right answer from a wrong one. Prices must not come from stale text, and a bad crawl must never wipe the search index.

What we built

1. A curated question-and-answer set, each entry tagged with how it may be served: as written, with a disclaimer, as a link only, held for review or hedged
2. A nightly Azure Functions job that reads the sitemap, crawls and cleans the pages, chunks and embeds the text and loads it into a hybrid keyword and vector index, with price masking, host allow-lists and an abort rule for a bad crawl
3. A chat engine in the storefront: fixed replies for greetings, contact, order status, price questions, abuse and self-harm, then retrieval with a stored answer returned only above set thresholds, and a model that answers only from the retrieved passages
4. An output guard that removes disease claims, quoted prices, leaked instructions, foreign links and emails
5. Safety-topic tracking across a conversation, with a caring reply and a crisis line for self-harm
6. Request limits, anti-forgery protection and a lazy-loaded, keyboard and screen-reader accessible chat widget that re-checks links before showing them
7. Automated tests for adversarial input, intent routing, the orchestrator, the output guard and live retrieval, and Bicep templates for the Azure resources
8. Removal of the previous bot and its in-browser answer data