Skip to content

Menu

RAG Development Your AI quoted last year’s price. The old file won.

We fix the files and the search.We fix the files and the search behind it, then test it with your team on real questions.

This form needs JavaScript.

Reach us another way:

WhatsApphello@supertoasty.com

Tell us about your documents

Our team will review your details and contact you about the next steps.

  • Old versions kept out of the answers
  • Tested with your team on real work

AI specialist

Why it went wrong

Wrong page in. Wrong answer out.

The AI writes a sure answer from whatever page the search brings back.

AB-2207 max load?

Exact part number

AB-2207 · Rev C · p. 4

Max load: 380 kg

The right page.

Close matches

  1. AB-2201 · 250 kgWrong part.
  2. AB-2270 · 410 kgWrong part.
  3. AB-2027 · 300 kgWrong part.
  4. AB-2207-OLD · 350 kgWrong part.
  5. BA-2207 · 200 kgWrong part.
A sample search
  • Codes need exact words. On exact-word searches, Microsoft scored keyword search 79.2 out of 100; meaning alone, 11.7.

  • Both kinds of search, then a second check. With labelled pieces: 67% fewer failed searches in Anthropic’s tests.

  • You see the page. We can show the document and page behind every answer.

Sources: Microsoft, 18 Sep 2023; Anthropic, 19 Sep 2024; checked 17 Nov 2026.

What you get

  • A map of your files
  • Cleaned files
  • Search by code and meaning
  • Answers in your tools
  • Tested on your questions
  • Monitoring after launch

Old versions

Old lists in a box. One list is live.

Nothing tells the search which one is live, so it picks whichever matches best.

  • One current version per topic. A named person on your side picks the live file.

  • Old versions kept, never searched. In the archive, out of every answer.

  • A check-by date on pages that go stale. Past its date, a price list flags itself.

Three price lists in your drive? Tell us where they live.

Get Started

Reading your files

12 months’ warranty. Which product?

The search reads files in short pieces; each one must carry its name.

  • Every piece keeps its document, product and section

  • Tables kept as tables, scans checked by a person

Do you need RAG?

Under 500 pages? Skip the build.

Anthropic, maker of Claude, advises putting up to about 500 pages straight into the AI’s prompt.

You need a build when

  • Thousands of pages, or files that change weekly
  • Answers turn on part numbers, codes, tables or scans
  • Answers belong inside your own product or website

You can skip it when

  • A few hundred tidy pages that rarely change

Source: Anthropic, 19 Sep 2024, checked 17 Nov 2026.

Testing

Your questions. Scored before launch.

Before anyone relies on it, it answers your expert’s questions and gets scored.

A sample test sheet: four questions, each marked for right page and right answer.
Your questionsRight page?Right answer?
Notice period in clause 14?YesYes
AB-2207 max load?YesYes
Warranty on the X200?YesNo
Do we ship to Brunei?Said: not in our filesno page neededYes
A sample sheet
  1. Your expert writes the questions. Real ones, with the right answers.

  2. Two scores for every answer. Right page and right answer; a true “not in our files” passes.

  3. The same test after every change. New files, settings or AI model: it runs again.

Your team

A team for your files. One name to call.

An AI specialist builds the AI that answers only from your own files.

AI specialist

Summarises calls, emails and documents.

After you press Submit

  1. 1We review your details.
  2. 2We talk it through with you.
  3. 3We propose a solution.
  • The systems, the code and the accounts are yours.
  • The tools run on your own accounts, so you can see what each one costs.
  • Priority access to Luna. Partners go to the front of the queue with our AI assistant.
  • We accept only a limited number of partnerships each quarter.

Questions

Can’t ChatGPT just read our files?

Cost, your data, messy files, and how you’ll know it works.

What is RAG development?

RAG development is building an AI that answers questions from a business’s own documents. RAG stands for retrieval-augmented generation: the AI first finds the right pages in manuals, contracts or price lists, then answers only from them. SuperToasty builds RAG systems for businesses with thousands of pages, files that change, or answers inside their own product.

  • RAG doesn’t train the AI on your files

A question we missed? Ask it in the form’s last box.

Get Started
Can’t we just upload our files to ChatGPT or Copilot?

Uploading files to ChatGPT or Copilot works well for a small, tidy set that rarely changes. Anthropic advises putting anything under about 500 pages straight into the prompt (19 Sep 2024, checked 17 Nov 2026). An uploaded copy stays as it was when the original changes, and you get little say in how files are searched. A RAG build earns its place with thousands of files, part numbers, scans, rules on who sees what, or answers inside your own product.

What does a RAG system cost to build and run?

The cost of a RAG system depends mostly on the files: how many, how messy, where they live and where the answers appear. The computing is cheap. At OpenAI’s published prices, turning 5,000 pages into searchable form costs about 7 US cents at about 500 words of text a page (our sum from OpenAI pricing, checked 17 Nov 2026). We put scope, price and timeline in writing before work starts, and the AI runs on your own account, so you see every charge.

Why does our AI quote old prices?

Old prices and policies usually come back because the search found an old file. A drive often holds several versions of the same price list, and nothing tells the search which one is current, so an old one can match the question best. The fix is in the files: one current version per topic, a named owner, and old versions kept out of the search.

Does our data train the AI?

A RAG system does not train the AI on your files: it looks the right pages up each time someone asks a question. The AI services we build on don’t train on them either: business data sent through OpenAI’s API is not used to train OpenAI’s models unless the customer opts in, and Anthropic’s commercial terms say it may not train models on customer content (both checked 17 Nov 2026).

How do we know it gives the right answers?

A RAG system should be proven on your own questions before anyone relies on it. Your expert writes real questions with the right answers, and each answer is scored twice: right page, right answer. Questions your files cannot answer go in too, and a true “not in our files” counts as a pass. We test with your team on real work.

RAG or fine-tuning: which do we need?

Fine-tuning changes how an AI model writes; RAG changes what it knows at the moment it answers. Prices, policies and specs that change belong in RAG, because updating the file is enough and each answer can point to its page. Fine-tuning suits a fixed style or format, and it does not keep facts current.

A question we missed? Ask it in the form’s last box.

Get Started

Keep going

Not quite what you need?

See the rest of this group, or the service closest to this one.