Scans in, text out
Pages are scanned, read by OCR and checked. Printed English is what it is built for; handwriting is flagged for a person.
For records and admin heads
Paper files · scanned PDFs · shared drives
Take the archive off paper. Then ask it questions.
Somewhere in the building there are cabinets of files that only one or two people can find their way around. Archive Search turns those files, and the folders of scanned PDFs on the shared drive, into a record that anyone with the right permission can search. Pages are scanned, read by OCR, checked and indexed; staff then ask in plain words and get an answer with the page it came from, shown next to the original scan.
scan_0417.pdf · page 1 of 2 · 300 dpi
Extracted fields
FieldValueConfidence
Document typeSupplier invoice0.97
SupplierSample Traders (Pvt) Ltd0.96
Invoice no.INV-2023-04170.98
Invoice date12 Mar 20230.95
PO referencePO-44710.93
Total (PKR)184,500.000.91
Approved byhandwritten Review0.62
Ask the archive
Question
Which invoices from Sample Traders were approved in March 2023?
AnswerTwo invoices from Sample Traders (Pvt) Ltd carry an approval stamp dated March 2023: INV-2023-0417 (PKR 184,500.00) and INV-2023-0432 (PKR 96,200.00). A third, INV-2023-0440, has no approval mark on the scanned copy.
Cited pages
scan_0417.pdf p.1scan_0432.pdf p.1scan_0440.pdf p.2
Grounded in 3 pages · Gemini on Vertex AI
What it does
Key fields pulled out
Dates, numbers, names and amounts are extracted from each page with a confidence score, and low scores go to a reviewer.
Find by number or meaning
An exact invoice number or CNIC, or a plain-words question about what a file says. Both work on the same archive.
Answers cite their page
The answer shows which page it came from, and the original scan opens beside it so a reader can check.
Your permissions, kept
Access follows your folder structure. A person finds only the files they could have opened in the cabinet.
One cabinet at a time
Start with one year or one department. Pages are indexed as they are scanned, so the archive is useful before it is complete.
How an Archive Search pilot works
A fixed-price Discovery first, then a fixed-price pilot on one set of files. The three stages are described on the homepage; this is what they mean for Archive Search.
-
You give
50 to 200 sample pages from one cabinet or one year of files, and the questions your staff actually ask about them.
-
We measure
How often the right page is found, the accuracy of each extracted field, and the time to find a file before and after.
-
You get
A searchable archive of those pages, an accuracy report, field by field, and a scanning plan for the rest.
Pilot price fixed in writing at the end of Discovery.
Google Cloud services it uses
| Service | Used for | Why |
|---|---|---|
| Cloud Storage | Original scans, versioned | Originals are kept as scanned, so every answer can show its source. |
| Document AI | OCR, layout and form parsing | Does the page work: text, tables and fields from each scan. |
| Vertex AI embeddings | The meaning of each page as a vector | Lets a plain-words question find pages that use different words. |
| AlloyDB for PostgreSQL | Keyword, vector and permission filter in one query | One store, one query; Vertex AI Vector Search for very large archives. |
| Gemini on Vertex AI | Answers that cite their pages | Answers only from the retrieved pages, with the citations shown. |
Most of the load is Document AI page processing, embedding generation on Vertex AI, and AlloyDB for retrieval; a single archive can bring hundreds of thousands of pages. The deployment is its own Google Cloud project, in the region you choose.
On your side: a connector in a container or VM that reads the shared drive, makes outbound TLS connections only, and can be switched off by your IT team at any moment.
Security and data notes
Your project, your region
Scanned pages and their text are stored in your own Google Cloud project, in the region you choose, and go to no service outside it.
Access from your folders
Permissions come from your directory and folder structure, and results are filtered before the model sees them.
No training on your data
Gemini on Vertex AI, under Google Cloud's published data-use terms, which do not allow Google to train its models on your documents without your permission. We do not train on client data either.
Questions records teams ask
Our archive is English, and some of it is handwritten.
Archive Search is built for printed English. We test OCR quality on your sample pages in Discovery before agreeing accuracy targets, and if the archive holds pages in other languages we look at those there too, without a target until we have seen them.
Handwriting is extracted where it is legible, and any field below the confidence threshold is flagged for a person. Search still covers the whole page, and every answer links to the scan itself, so a reader can check.
Do we have to scan everything before we start?
No. The pilot starts with 50 to 200 sample pages from one cabinet or one year of files, and the archive grows from there. Scanning can be done by your staff or a scanning bureau at the resolution and format we agree in Discovery, and pages are indexed as they arrive.
Where are the scanned pages stored?
In Cloud Storage, with versioning, inside a Google Cloud project dedicated to you in the region you choose. We operate that project for you, or it sits inside your own Google Cloud organisation if you prefer to own it.
Who can see what?
The same people who could open the folder today. Permissions are mapped from your directory and folder structure in Discovery, and every search is filtered by them before any answer is written.
Talk to us about Archive Search
Tell us roughly how many cabinets or boxes, what kind of records, and who needs to find them. The form is on the homepage, or write to contact@technometa.com.
Go to the contact form