AI document reading
The model reads the page. Code does the maths.
Audited accounts, agreements and org charts are read by an AI model from the page images, through TAIGA’s own model gateway. Each figure comes back as printed, with the page it came from, and every calculation after that is code.
What it reads
What goes in, and how it is read.
Word and PowerPoint files can be uploaded, but they are not read yet. Ask your client for a PDF export.
- PDFRead as page images
- ImagesRead as page images
- Excel workbooksRead as tables
- Word and PowerPointExport to PDF first
Only the statements
A 148-page annual report, read where it matters.
On a report over 50 pages, TAIGA first finds where the standalone statements start and end, then reads its figures from that range only. The consolidated set is not mixed in.
- Up to 100 pages are read in one pass.
- Files up to 100 MB are accepted.
Annual report
148 pages
- 1FindLocate where the standalone statements start and end.
- 2Pad3 pages either side absorb printed page numbers that drift from the PDF.
- 3ReadOnly that range is sent to be read.
Evidence, not a summary
Every figure quotes its section and page.
The extraction prompts require each value to arrive with a citation: the section, the page, then the line as printed. TAIGA shows that quote exactly as written, never tidied. A citation without a section and page is not shown as one.
- Figures arrive in the form marked as read from a document.
- Correct anything that looks wrong. Saving the step confirms the rest.
[Audited P&L p.8] Revenue 264,202,244
- Where
- Audited P&L p.8
- As printed
- Revenue 264,202,244
[Note 16 p.27] Provision for employees’ end of service benefits 8,009,425
- Where
- Note 16 p.27
- As printed
- Provision for employees’ end of service benefits 8,009,425
validation check
No section and page, so it is not shown as a citation.
Lakhs and crores
The model reads the unit. Code applies it.
The model returns each figure exactly as printed and names the statement’s unit once. Code multiplies, so a lakhs statement can’t be read a hundred thousand times too small, or scaled twice.
- Related-party notes printed in full amounts inside a lakhs statement are detected and left unscaled.
The model returns
2,642.02
Statement unit: lakhs
Code stores
264,202,000
× 100,000
| Statement unit | Multiplied by |
|---|---|
| Thousands | 1,000 |
| Lakhs | 100,000 |
| Crores | 10,000,000 |
| Millions | 1,000,000 |
| Billions | 1,000,000,000 |
While it reads
The wait says what it is reading.
After upload, your client sees how many documents are read out of how many, and the latest figure found with its citation. Failed documents and documents they stopped are counted separately.
- A read in progress can be stopped.
- If it takes too long, keep waiting or go back to the documents. Nothing moves on without the read.
Where the model stops
It reads numbers. It never computes one.
Totals, operating revenue and operating cost are derived by code from the lines read, even when the model returns its own. The PLI, the range and where the tested party sits are ordinary arithmetic you can check.
The model
- Reads each figure as printed
- Names the statement unit
- Cites the section and page
- Sorts lines into income and expense items
Code
- Scales figures to absolute amounts
- Derives totals, operating revenue and cost
- Computes the PLI
- Computes the range and the tested party’s position
Coverage
What it covers.
- Formats read
- PDF, images, Excel
- Word and PowerPoint need a PDF export
- Pages per read
- Up to 100
- Statements located first on long reports
- File size
- Up to 100 MB
- Per upload
- Reporting units
- Five scales
- Thousands, lakhs, crores, millions, billions
- Standards
- Ind AS and IFRS
- A reading prompt for each
- Logs
- Metadata only
- No document or prompt content is logged
Questions
Before you ask.
- Can my client upload Word or PowerPoint files?
- They can upload them, but they are not read yet. Ask for a PDF export. PDFs and images are read from the page, and Excel workbooks are read as tables.
- What if it misreads a figure?
- Each figure shows the section and page it came from, so you can check it against the PDF. Correct it in the form, and your figure is the one used.
- Does the model calculate the PLI?
- No. The model reads figures. The PLI, the range and the tested party’s position are calculated by code.
- Can we see which prompt read a figure?
- Yes. Every extracted value records the model and the prompt version that produced it.
