Store
The tool tells people when it can’t find an answer, and questions with only weak matches are logged.
How we build it
Every change to this tool is approved before it starts, planned before it's coded, checked by a second AI, and proven before it goes live. Every step is recorded.
Try the Tradify answer tool →
What RAG is
RAG stands for retrieval-augmented generation. Instead of asking an AI to answer from its general memory, we first retrieve the right passages from your own documents, then have the AI write the answer from them.
It works like an expert who checks the manual before answering, then shows you the page.
RAG explained · a beginner's guide (PDF)
New to RAG? Ten slides, plain English, using Tradify as the example.
The answer path
1 · Ingestion
2 · The special database
Vector databaseEvery passage is stored with numbers that capture its meaning.passage 1 → [0.12, −0.48, 0.91 …]
passage 2 → [0.07, 0.33, −0.25 …]
passage 3 → [−0.61, 0.18, 0.44 …]2,631 passages live
● Chat session
How do I turn an accepted quote into a job?
Identify the gaps
When a question can't be answered well, that's useful information: it points to an article your help centre doesn't have yet, or one that needs rewriting.
The tool tells people when it can’t find an answer, and questions with only weak matches are logged.
Every unanswered question kept in people’s own words, with similar ones grouped so you see the topic and how often it comes up.
A weekly report with a suggested article title and outline for each group, for your help desk team to write.
By the numbers
Article and passage counts are live. The cost covers the AI model for each answer; hosting and the database are a fixed monthly cost, not per question.
| 826 | help articles indexed |
|---|---|
| 2,631 | searchable passages |
| 1,000+ | questions answered |
| under 20c | AI cost per question |
The answers
The tool is built to answer only from your help-centre articles. It finds the most relevant passages, writes a step-by-step answer from them, and links the articles it drew on, so you can check it.
If the articles don't cover a question, it's built to say so instead of guessing.
The stack
Vercel for the product experience, Supabase for state, n8n for workflows, and Cloudflare at the front door.
A full build record
Since October 2026, every Issue, plan, review and change for this tool is recorded. Ask us and we'll walk you through it.
Ask for a walkthrough →For your business
Talk to us about having your own knowledge base tool, built on your help documents.