AI Chatbot for Chroma Data,
Natural language data access across multi-cloud infrastructure
AI Chatbot for Chroma Data — AI Case Study
A conversational AI chatbot enabling natural language queries against enterprise data, with seamless Microsoft Teams integration and a secure multi-cloud architecture for in-app data access and insights.
Three numbers that
tell the story upfront.
Documents indexed
A conversational AI chatbot enabling natural language queries against enterprise data, with seamless Microsoft Teams int…
Median query response time
A conversational AI chatbot enabling natural language queries against enterprise data, with seamless Microsoft Teams int…
Answer accuracy on eval set
A conversational AI chatbot enabling natural language queries against enterprise data, with seamless Microsoft Teams int…
Before we got here,
here's what wasn't working.
The AI Chatbot for Chroma Data makes enterprise data effortlessly accessible through natural conversation. Built on a secure multi-cloud architecture spanning Azure and AWS, it uses an AI agent with tool-calling capabilities to run database queries and return insights in plain English. The chatbot integrates seamlessly with Microsoft Teams, so teams can query and analyse their data without leaving their day-to-day workspace.
We confirmed every challenge with the team before writing a line of production code. The right column lists the pain points we tackled head-on and how we approached each one.
What wasn't working
BEFOREChunking strategy for documents with tables and charts
Used a hybrid approach: text chunks for prose, separate table extraction via Camelot for structured data, stored as distinct metadata-tagged embeddings.
Don't over-engineer.
Solve what matters, ship fast.
The instinct on complex projects is to rebuild everything from scratch. We pushed back. A full rewrite often means months of regressions while engineering rebuilds features that were already working.
We mapped the existing workflow against actual usage. The result was clear: a small set of features accounted for the majority of user time. We rebuilt the core, shipped fast, and extended outward.
On the technical side, we made clear decisions early and held them. The right column shows what we chose, and why.
4 decisions we made early
CHOSEN / REJECTEDNatural Language Data Queries — how we built it
Custom Natural Language Data Queries — built for this product's exact needs and data model.
Generic SaaS tool that would require heavy workarounds and compromise the user experience.
AI Agent with Tool Calling — how we built it
Custom AI Agent with Tool Calling — built for this product's exact needs and data model.
Generic SaaS tool that would require heavy workarounds and compromise the user experience.
Microsoft Teams Integration — how we built it
Custom Microsoft Teams Integration — built for this product's exact needs and data model.
Generic SaaS tool that would require heavy workarounds and compromise the user experience.
Secure Multi-Cloud Architecture — how we built it
Custom Secure Multi-Cloud Architecture — built for this product's exact needs and data model.
Generic SaaS tool that would require heavy workarounds and compromise the user experience.
4 features that
define the product experience.
Every feature was scoped and shipped with the same priorities: speed, reliability, and staying out of the user's way.
Natural Language Data Queries
Enables natural language conversations to retrieve and analyse data effortlessly, without needing technical knowledge.
AI Agent with Tool Calling
Utilises an AI agent with tool-calling capabilities to run database queries and return accurate, structured results.
Microsoft Teams Integration
Integrates seamlessly with Microsoft Teams for in-app data access and insights within the team's existing workflow.
Secure Multi-Cloud Architecture
Built on a secure, multi-cloud architecture leveraging both Azure and AWS for reliability, scalability, and data security.
What changed,
measured against the day-one baseline.
We benchmark before we touch anything, logging every metric below against its state before we wrote a line of production code.
Related services
Want to build
something like this?
We ship fast. Talk to us about your product idea and we'll tell you exactly how we'd approach it.