Document Extractor
Document Extractor turns a pile of PDFs into tables. Add the folders that hold your invoices, vouchers and bank statements, press Extract all, and it writes CSV and Excel files you can open, sort and send to your accountant. Every amount is checked against the page, and anything doubtful is marked for you to look at. Your PDFs are never changed.
Where to find it
Section titled “Where to find it”Open Document Extractor in the sidebar, under TOOLS. There is no switch to turn on. It uses the ZenithAI key you entered at first-run setup. If the key is missing, the page says “Add your ZenithAI key in Settings first.” If ZenithAI rejects it, the page says “ZenithAI rejected the key. Check it in Settings.”
Add your PDFs
Section titled “Add your PDFs”- Press Add folder… to choose a folder of PDFs. Use the arrow beside it, then Add PDF files…, to add single PDFs.
- Or drag folders and PDFs onto the page. While you drag, the page says “Drop to add” the folders and PDFs you hold.
- Add as many folders as you like. A PDF reached through two folders is listed once. To take a folder or PDF you added off the list, select it and press Remove, or right-click it and choose Remove from the list. Nothing on disk is changed.
Folders are searched for PDF files, and nothing else in them is read. Include subfolders is ticked by default. Folders whose names start with a dot, system and program folders such as Windows and Program Files, build and dependency folders such as build, bin and vendor, links, online-only files and the run’s own output folder are left out.
Choose what to read
Section titled “Choose what to read”The row under the buttons holds the choices for every run.
| Choice | What it does |
|---|---|
| Type | Auto-detect tells invoices, vouchers and bank statements apart. Pick Invoice, Voucher or Bank statement when every PDF is of that kind. A file whose type cannot be told is skipped with “Could not tell the document type; pick a type and re-run.” |
| Model | Genius (most accurate), Smart or Fast. Genius is the default. The page notes that faster models can misread values, and that every value is still checked. |
| Change… | Chooses where the files are saved. The folder you choose is remembered. Without a choice, they go to an Extracted folder inside the first folder you added, or, when the first thing you added is a PDF, inside the folder that holds it. |
| More | Opens Ignore saved results, Match with a ledger CSV…, Open output folder and What it has learned. |
If the output folder cannot be written, the app saves to its own exports folder instead and says so.
Run it
Section titled “Run it”Press Extract all. To read part of the list, select a folder or a PDF and press Extract this folder or Extract this file. Their files are saved in a folder of their own inside the output folder, named after that folder or PDF, so they never overwrite the files of the whole list. Stop ends the run now and keeps what was finished.
While it runs, a status card shows the current step and a progress bar, with six numbers: Files, Pages, Elapsed, Per file, Per page and Time left, about. When ZenithAI takes only a few requests a minute, the page says “Waiting for ZenithAI…” and that the pause is normal. If ZenithAI does not answer, it says so and tries again.
The page shows your folders as a tree on the left, each with a tally of files done, needing review, failed, in progress and skipped. Select a folder to see its files in a table with these columns: File, Type, Pages, Status, Time, Found and Notes. A narrow window shows fewer columns.
A file’s status is one of Ready, Queued, Reading pages, Extracting, Waiting, Checking, Done, Needs review, Failed, Skipped, Duplicate, Stopped or Not started. A failed or skipped file says why and what to do, for example “Remove the password and re-run”, “Split it into smaller PDFs” or “Press Extract again; finished files are kept”. Finished files are kept, so pressing Extract again with the same type and model uses the saved results and sends nothing for them. Another model or type reads them again.
Select one PDF to see its page on the left and, beside it, what was read from it. The arrows move between pages. Open PDF opens it in your PDF viewer. Right-click a file for Show what was found, Open PDF and Show in folder.
When the run ends, a banner gives the result in one line, for example how many documents and rows were found and how many files need review. Open folder opens the output folder. Open in Excel lists the Excel files of this run.
The files it writes
Section titled “The files it writes”Each file is written as a CSV and as an Excel workbook with the same name. A file appears only when the run found rows for it.
| File name | What it holds |
|---|---|
invoices |
One row per invoice: number, dates, supplier and buyer with GSTIN, taxes, totals, and the review notes |
invoice_lines |
One row per invoice line: description, quantity, rate, tax rate and amount |
vouchers |
One row per voucher: type, number, date, party, amount, payment mode and narration |
bank_statements |
One row per statement: bank, account, period, opening and closing balance, and the balance check |
bank_transactions |
One row per bank transaction: dates, narration, reference, debit, credit and running balance |
reconciliation |
Bank rows matched with your vouchers and invoices. Written when the run has bank rows and vouchers or invoices. |
reconciliation_ledger |
Bank rows matched with a ledger CSV you chose. Written only when you chose one and the run has bank rows. |
run_report |
One row per PDF with its status, documents found, rows written and the time it took |
In the reconciliation files, a bank row is matched, probable, ambiguous or unmatched_bank. A record with no bank row is unmatched_ledger. Rows are paired by the same amount to the paisa and a date within 7 days, or, for an invoice, up to 90 days after its date. A payment voucher pairs only with a debit and a receipt voucher only with a credit. When the number or reference of the voucher, invoice or ledger row is found in the bank text, the row is matched. A single possible partner without one is probable. When several partners are possible, or two bank rows could take the same one, the row is ambiguous, the possible partners are listed in candidates, and nothing is paired.
To match against your own books, open More, press Match with a ledger CSV… and choose a file exported from your accounts software. Its columns are found by name. A ledger is refused if it has no date or no amount column, or is over 10 MB or 50,000 rows. Clear stops using it.
Rows that need review say needs_review = yes and give the reason in review_notes. When you review a row, the user_review column records what you did, so the files keep the trail.
Excel details
Section titled “Excel details”- Amounts show two decimals.
- Dates are real dates, shown as year-month-day, for example 2026-04-18.
- Columns are sized to fit their contents, and the header row is bold and frozen.
- Identifiers such as account numbers and invoice numbers are stored as text, so leading zeros stay.
- The CSV files open in Excel with Hindi text showing correctly.
If you stop a run, or it ends before every file is read, the files are saved with -partial in their names, for example invoices-partial.xlsx. A complete set from an earlier run is never overwritten by them. If a file is open in Excel, the new one is saved as the same name with (2), (3) and so on.
The checks, and Needs review
Section titled “The checks, and Needs review”Every value is checked on your PC. A check only adds a note. It never changes a value.
- Totals: the lines add up to the subtotal, and the parts add up to the grand total.
- Line amounts: quantity times rate, less any discount, gives the line amount.
- Running balances: each bank row’s balance follows from the one before it, and the opening balance plus credits, less debits, gives the closing balance.
- GSTINs, PANs and IFSC codes: their form is checked, including the last character of a GSTIN. A PAN is compared with the supplier’s GSTIN.
- Dates: a date that cannot be read, that looks far in the past or the future, or whose day and month could be swapped is flagged.
- Values on the page: a value that is not found in the text read from the page is flagged.
- Duplicates: an invoice with the same number from the same supplier as an earlier one is flagged, and so is a voucher with the same number, date and amount as an earlier one.
A second copy of a bank statement already read in the same run, for example a scan of a digital one, is recognised and marked “Same statement as an earlier file; its rows are left out of the matching”. The rows stay in bank_transactions, but they are kept out of the reconciliation.
A flagged file shows Needs review with the reason in plain words, for example “The total does not add up” or “A GSTIN is not valid”.
Review a file
Section titled “Review a file”Click the Needs review status of a file, or select it and press Review. A folder’s Review button goes through the flagged files one by one, and Review next moves on to the next file that needs a look.
- The page opens on the left with each doubtful value boxed in yellow where it is found. On a digital PDF the box sits exactly on the value. On a scan the box is dashed and the page says “Dashed boxes: about here, found by ZenithAI on the scan”. Zoom shows the page nearer, centred on the active value.
- On the right, one card for each issue explains it in plain words and shows the value in a box.
- If the value is wrong, type the right one and press Correct it. You can also press the Use button when a suggestion is offered.
- If the value is right as it is, press It’s right. When there is no single value to correct, for example a page that could not be read, a balance that does not add up or a duplicate, the button reads Mark as checked.
- The CSV and Excel files are updated at once. The next open card becomes active. When none are left, the page says “All checked”.
Your corrections are kept for each PDF and applied again on every re-run, whether it reads the file again or not. If a later read gives another value than the one you corrected, the row says so and asks you to check it again. A scanned page needs your ZenithAI key to be marked.
Remember this
Section titled “Remember this”Under a card, a box reading Remember this for future invoices from the supplier’s name (or vouchers, or statements) is ticked by default. It is shown only when what you are about to do can be remembered for that supplier. When you correct a value, the next documents from that supplier are read with a hint. When you press It’s right, that check is no longer raised for that supplier. Sums, balances, duplicates and failed reads are never switched off, whatever you decided before.
To see or remove what it learned, open More and press What it has learned. The list shows each supplier with its Hint and Not raised entries, each with Remove.
Scans, digital PDFs and languages
Section titled “Scans, digital PDFs and languages”Scanned and digital PDFs both work, in English and Hindi. Hindi digits are read as numbers. Amounts written the Indian, Western or European way are understood. Your PDFs are only read, never saved, moved or deleted.
What goes to ZenithAI
Section titled “What goes to ZenithAI”The reading is done by ZenithAI, with the key you already use for Chat.
- When you press Extract: the pages of each PDF go to ZenithAI to be read, under a made-up file name rather than the real one. The page pictures and the text read from them go with them, with the instruction to find the values. A very large file is sent as pictures of its pages. Nothing is sent when you only add files.
- When you correct a value with Remember this ticked (the default): a short note about that supplier, such as how its invoice number is printed, goes with that supplier’s later documents when they are read.
- When you review a scan: each scanned page that has a value to check goes to ZenithAI as a picture, with the values to find on it, once, when you reach it. The answer is kept. If the call fails, it is asked again the next time you open the file. On a digital PDF, the box is found on your PC and nothing is sent.
- Stays on your PC: your PDFs, the checks, the CSV and Excel files, your corrections, the list of what it learned and the saved reads. The saved reads let a re-run skip files already finished. They hold document data, as the CSV files do, and are plain local files. See How your data is handled.
Limits
Section titled “Limits”- A PDF can have up to 500 pages. A larger one says “Split it into smaller PDFs”.
- A password-protected PDF is not read. Remove the password and run again.
- A file that is damaged, empty or not really a PDF is marked Failed with the reason. A file open in another app says to close it and press Extract again.
- ZenithAI takes only a few reading requests a minute, so a big folder takes time and the page shows “Waiting for ZenithAI…”.
- If ZenithAI does not answer three requests in a row, the run stops, finished files are kept, and the page says “ZenithAI could not be reached. Check your internet connection, then press Extract again; finished files are kept.”
- Faster models can misread values. Every value is still checked, and doubtful ones are marked.
- Dropping folders onto the page does nothing while a run is active.
Troubleshooting
Section titled “Troubleshooting”| What you see | What to do |
|---|---|
| “Add your ZenithAI key in Settings first.” | Enter your key in first-run setup or under Settings > General > Change API key…, then press Extract again. |
| “ZenithAI rejected the key. Check it in Settings.” | The key was refused. Check it in Settings > General. |
| “The ZenithAI address in Settings cannot be used. Check it, then press Extract again.” | The service address saved at setup is not valid. Reopen setup with Edit profile… in Settings > General and check it. |
| “ZenithAI could not be reached.” | Check your internet connection and press Extract again. Finished files are kept. To see why, press Check connection in Settings > General. The FAQ explains its messages. |
| “ZenithAI was too busy. Wait a minute.” | Wait a minute, then press Extract again. |
| “Reading the pages took too long.” | Press Extract again. Finished files are kept. |
| A file is Failed or Skipped | Read its note in the Notes column, or select the file. The note says whether to remove a password, split the file, pick a type or run again. |
The files are called -partial |
The run did not read every file. Press Extract again to write a complete set. |
| An Excel file did not update | Close it in Excel. If it was open, the new file has (2) in its name. |
Related pages
Section titled “Related pages”Common questions
What does Document Extractor do?
It reads invoices, vouchers and bank statements from PDFs, scanned or digital, in English or Hindi. It saves what it read as CSV and Excel files in an Extracted folder inside the first folder you add. Your PDFs are never changed.
What does Document Extractor send out?
The pages of the PDFs you extract go to the ZenithAI service, which does the reading, with your existing key and under a made-up file name. A correction kept with Remember this sends a short note with that supplier's later documents. Reviewing a scan sends its pages with values to check. The checks and your results files stay on your PC.
What does Needs review mean?
A check found something doubtful: a total that does not add up, a running balance that breaks, an invalid GSTIN, an unclear date or a missing value. The file is marked with the reason. The value is never changed for you.
How do I fix a value marked Needs review?
Click its Needs review status. The review screen shows the page with each doubtful value it finds boxed in yellow. Type the right value and press Correct it, or press It's right. The CSV and Excel files update at once, and your correction is kept on every re-run.
Does it work on scanned PDFs and Hindi documents?
Yes. It reads scanned and digital PDFs, in English and Hindi. On a digital PDF the review boxes sit exactly on the value. On a scan they are drawn dashed and say about here, because ZenithAI finds the place.
Can I run it again without losing my corrections?
Yes. Corrections are kept per PDF and applied again on every re-run. A stopped run saves files with partial in their names and never overwrites a complete set from an earlier run.
Does the Extracted folder contain my original PDFs?
No. The PDFs stay where they are and are only read, never saved, moved or deleted. The Extracted folder holds only the CSV and Excel files the tool writes.
