How we gave 200 staff answers from their own manuals, over 95% right
We put an assistant in the staff portal for 200 people, with over 95% correct answers on questions written with the client’s team.
- Gemini
- Vertex AI
Gemini
- Client
- Steel detailing firm of 200 staff, serving North American fabricators
- Documents
- Training courses, software guides and internal instructions
- What we built
- A manual reader, a search and an assistant on the firm’s Google Cloud
- Where it runs
- A chat page inside the staff portal they already use
- Live since
- May 2026, with a second phase in July
GeminiA library of manuals
Thousands of pages of software guides and in-house standards. Much of what staff need sits in screenshots and tables that a search box can’t read.
Each page read as needed
The system checks each page for pictures. It reads screenshots and tables in detail and uses a cheaper reading for plain text. This cut the reading cost by 42% on the first 1,581 pages.
Tables recovered from screenshots
The detailed reading copies each screenshot’s rows into a table and adds a description. This gives the search enough information to find the right row.
Answers with a source
Staff ask a question and get an answer with a link to the relevant row in the manual. Over 95% of answers were right on questions written with the client’s team. Staff check the source before using an answer.
One question, asked four ways.
Staff don’t phrase questions the way manuals are written.
Staff question
What’s the prefix for angle steel?
Rewritten four ways
- 1angle steel prefix
- 2naming code for L sections
- 3letter that marks angle profiles
- 4equal angle section designation
Merged, duplicates removedfound by
- 1Naming standard, section 4.24 of 4
- 2Section table, page 123 of 4
- 3Drawing notes, 2.12 of 4
- 4Shop guide, section 72 of 4
- 5QA checklist, item 31 of 4
Answer
L, per the naming standard
Source: naming standard, section 4.2
+
answers correct on the client’s test questions
asked one way: 75%
- 95%+correct on questions written with the client’s team, up from 75%
- 42%lower reading cost, reading only picture pages carefully
- 40%faster replies after tuning the hosting
- 200staff with answers from their own manuals
“The chat assistant is very good. Please launch it on the live site.”
Read the full story
A structural steel detailing firm kept its working instructions in thousands of manual pages. Many answers were inside screenshots that text searches missed. We built a system that reads those screenshots and answers staff questions inside their existing portal.
The answer is in a screenshot that search can’t read.
Detailing runs on software, and the know-how for that software lives in manuals full of screenshots, tables and drawings.
New detailers use the training courses, and experienced staff refer back to the manuals. Often the instruction they need sits in a screenshot of a software window. A search of the page’s text misses it.
The firm already paid for the cloud services it needed. We had to make the screenshots searchable and help the assistant find the relevant page when staff used different words from the manual.
“If it gives rubbish answers, the staff will test it and decide it has no value.”
Every screenshot turned into words.
We built a reader that treats each page according to what is on it.
The system checks each page for pictures. Screenshots and drawings get a detailed reading that produces tables and written descriptions. Plain text pages use a simpler reading. Sorting the pages this way cut the reading cost by 42%.
We also stripped out repeated logos so they never clutter an answer, tagged each manual with the software it covers, and kept every image in private storage, linked to the exact passage it belongs to. Thousands of pages can run through without starting over if anything stops halfway.
“The way it turns screenshots into tables of data is amazing. It is far better than when we put the documents into Gemini ourselves, without these added descriptions.”
One question, asked four ways.
Staff don’t phrase questions the way manuals are written.
We put the processed pages into a search library on the firm’s own cloud, hosted in Europe. The assistant can rephrase a question three more ways, search all four versions and combine the best ten passages. It answers from those passages, and staff can read the reply as it appears.
On questions written with the client’s team, that approach lifted correct answers from 75% to over 95%. A first step also works out which set of manuals a question belongs to, so each search looks in the right place.
Live for every member of staff.
Staff could use the assistant through their existing portal login.
The assistant sits on a chat page inside the staff portal the firm already used. A dashboard shows who asked what, how fast it answered and what it cost, with an export for the managers.
We released the manuals in batches, and the client’s own team tested each batch before it went live. The assistant went live for all staff in May, and a second phase added thousands more pages in July.
Most questions turned out to be short lookups, so we switched from the most capable model to faster, cheaper models. We also tuned the hosting with the client’s team. Replies came back 40% faster.
Try the questions your staff keep asking.
Send us the manuals your team keeps asking about. We’ll show you the assistant answering from them.