Thousands of pages of content and years of forum history were overwhelming visitors and the marketing team, who were answering the same questions over and over.
The Problem
Our client, a large yoga and wellness retreat brand across Australia and Thailand, had grown into a complex business: a hotel, a cafe, and yoga offerings spanning day visitors taking a single lesson, retreat guests staying for a program, and long-term students training to become instructors. Their content spanned 2,000 to 3,000 pages, covering dozens of yoga styles that differ in small but important ways, plus a community forum built up over five to ten years. Visitors struggled to find what they needed, and the small marketing team was buried answering repetitive questions about pricing and discounts, slowing down conversions.
What We Built
We built an AI chatbot that actually understands the retreat’s full offering, not just pre-written answers, and can guide a visitor to the right information instantly, whether they’re booking a single class or signing up to become a teacher.
Getting this right took more than a standard approach. A regular language model fed the website’s content gave answers that were close but not precise enough, and when yoga styles and course tiers differ by small details, close isn’t good enough. We combined vector search with a graph database (Graphiti) into our RAG pipeline, so the system could understand how offerings and pricing actually related to each other, not just match keywords.
We also pulled in years of forum history, so the chatbot could draw on real community answers alongside official product pages. We connected the chatbot to the client’s CRM, so every conversation and lead gets logged automatically, and built in promotion awareness, so the system surfaces the right offer when a visitor shows interest. When a question needs a human, the AI hands the conversation off smoothly.
The Result
- Support ticket volume dropped by 28% in the first two weeks
- The AI answered the majority of repetitive questions without staff involvement
- Customer experience improved dramatically, since visitors no longer had to dig through thousands of pages
- The sales team gained complete visibility into every interaction, automatically recorded in HubSpot
- The system scaled cleanly across multiple campuses and course types, supporting the brand’s growth
The Honest Bit
The hardest part wasn’t the chatbot itself; it was making the underlying data trustworthy. With 2,000 to 3,000 pages covering yoga styles that differ in very small ways, our first approach, a standard model, gave answers that were close but not precise enough. When someone’s asking exactly which discount applies to which course, close isn’t good enough.
After real trial and error, combining vector search with a graph database gave the system enough understanding of how things related to each other to consistently pull the exact right answer, regardless of phrasing. That precision is what moved the support ticket numbers.
The other genuinely hard problem was the community forum, built up over five to ten years, that the client wanted folded in too. The same question had often been answered differently over time, as pricing and programs changed. The system had to recognize that and always serve the current, correct answer, not the most recent-sounding one, even when an older answer technically matched. Getting that right mattered, since surfacing the wrong historical answer would have been worse than not answering at all.
Book your AI strategy call
Please schedule a AI consultation call. We’ll discuss your business needs and explore how AI can enhance your operations and drive growth.
