AI support assistant for a SaaS platform
An assistant that answers customer questions from the help center and product docs, cites its sources, and hands off to a person with the full context when it cannot help.
- Client
- B2B project management software company
- Location
- Australia
- Platforms
- In-app widget, help center, support inbox
- Duration
- 10 weeks
- Team
- 3 engineers
The challenge
Support volume was growing faster than the team. Most tickets asked questions already answered in hundreds of help articles, but keyword search missed them, and an earlier chatbot pilot gave confident wrong answers.
What we built
We built a retrieval-augmented assistant over the help center, release notes and API docs. Answers are grounded in retrieved passages and link to their sources. An evaluation set built from real tickets runs on every change, so prompt and model updates are measured before release. Conversations the assistant cannot resolve become tickets with a summary attached.
Highlights
- Hybrid keyword and vector search over the docs, re-indexed on publish
- Every answer cites the articles it is based on
- Evaluation suite of past tickets that gates each release
- Hand-off to the support inbox with a conversation summary
- Dashboards for usage, resolution rate and feedback
Technical stack
- AI
- Claude APIEmbeddingspgvector
- Backend
- PythonFastAPIPostgreSQL
- Frontend
- ReactTypeScript
- Infrastructure
- Google Cloud RunGitHub Actions
Have something to build? Let's talk it through.
Tell us what you are working on. We reply within one business day, and after a short discovery call you get a written scope and estimate, free of charge.