A cited AI assistant for Orbiill's billing platform.
How we took Orbiill's AI assistant from proof of concept to production: answers from Orbiill's own docs, a source behind every answer, and accuracy re-tested before every change.
Illustrative recreation · no client data
Illustrative recreation of the production interface — no client data shown. The assistant answers inside the pricing editor, with its sources.
The challenge
Orbiill helps SaaS teams configure pricing and billing on top of Stripe. Its users have questions about plans, quotas, usage-based overage and deploying to Stripe — and the answers live across docs, FAQs and plan pages.
Orbiill wanted an assistant inside the product that answers those questions correctly and shows where each answer comes from. The starting point was a proof of concept; the job was to make it accurate enough for customers, measurable, and part of the product.
What we built
An assistant that lives inside the product and answers from Orbiill's own content, with sources.
- Finds the right answer. It searches two ways at once — by exact words (plan names, limits, Stripe settings) and by meaning (questions in the user's own words) — then ranks the best passages before answering.
- Knows the customer. Answers take into account each customer's billing setup and the conversation so far, so follow-up questions stay on topic.
- Shows its sources. Every answer links to the documents it came from, so users can check it and trust it.
- Part of the product. Built into Orbiill's app, next to the pricing editor — not a separate chatbot page.
How it works under the hood
Hybrid retrieval (BM25 keyword search plus vector search on pgvector), fused and reranked with Cohere, then answered by Claude with citations. Every change to chunking, retrieval or prompts re-runs the eval set before it ships.
How we measured quality
Before tuning anything, we wrote an eval set (a fixed list of test questions): real questions Orbiill's users ask, each paired with the source that should answer it. Every change to how the assistant searches or answers was scored against it, and only changes that held up shipped.
This is the core of how we work: quality is a number you can watch, not a feeling after a demo. The same eval set stays with the client, so their team can keep improving the assistant safely after handoff.
What shipped and what the client said
The assistant went from proof of concept to production and was integrated into Orbiill's product, answering with source citations and context about each customer's setup.
“Working with Julian Lopez to develop the AI Chat features for orbiill.com was a seamless experience. He successfully implemented a powerful RAG system that has significantly leveled up our user experience. Professional, highly skilled, and efficient — highly recommended!”
Stack (for the technical readers)
Want the same for your product?
If your users have questions your docs already answer, we can build an assistant like this on your content — scoped, priced and measured up front. Most RAG builds take 3–5 weeks.