Can AI pull tables out of PDFs? Yes. The right tool depends on whether the PDF has text and ruled lines.
Short answer. Yes, and the right tool depends on the PDF. If you can select the text in a PDF viewer, rule-based libraries such as Camelot and pdfplumber read the tables directly, faster and more predictably than a language model. If the PDF is a scan, you need OCR first, then a table-structure model such as Microsoft’s Table Transformer or a document converter such as Docling. Tables without ruled lines, merged cells and tables that run across pages are the usual failures. Whatever you use, reconcile the output: row and column totals must match the printed ones.
01
Which tool for which PDF
Tool
Best for
Licence
Limits stated in its documentation
pdfplumber
Tables in machine-generated PDFs, with detailed access to every character and line.
MIT
Its README says it works best on machine-generated, rather than scanned, PDFs.
Camelot
Ruled tables (lattice parser) and whitespace-separated tables (stream parser).
MIT
The built-in parsers need a text-based PDF. Scanned or image-only PDFs need its optional machine-learning and OCR backend.
Table Transformer, Microsoft
Finding tables and their rows and columns in page images.
MIT
Trained on PubTables-1M, nearly a million tables from scientific articles, so test it on your own documents.
Docling
Converting PDF, DOCX, PPTX and XLSX files with table structure and OCR into structured output.
MIT
Its README highlights local execution for sensitive data and air-gapped environments.
Vision-language models, for example Qwen2.5-VL
Reading a table from a page image into structured output.
Apache 2.0
Its card says it supports structured outputs for scans of invoices, forms and tables. Check the cell count against the page.
02
Why tables are hard
A PDF stores where each character sits on the page, not that a group of characters forms a row. Rule-based tools rebuild rows and columns from positions and ruling lines, which works well when the lines are there. Camelot offers separate parsers for ruled tables and for tables separated only by whitespace for that reason. A model given only the extracted text has lost the positions, and a vision model has to infer the grid from the picture.
In practice the trouble spots are borderless tables, merged header cells, one table split across pages with repeated headers, rotated pages, and scans that are skewed or faint.
03
What a person must still check
Totals. Row sums and column sums in the extracted table must equal the printed totals.
Column alignment. A shifted column, where values slide into the neighbouring field, is the most common silent error.
Merged cells and spanning headers: does every value sit under the right heading?
Units and scale in headers, such as ₹ lakh and ₹ crore, and numbers in Indian digit grouping like 1,23,456.
Negative numbers in brackets, dashes meaning zero, and blank cells meaning not applicable.
Tables continued across pages. Count the rows against the source.
04
How an office would set this up
Sort the PDFs first: text-based, scanned, or mixed. The select-a-line test in a PDF viewer is enough.
Use a rule-based tool for text-based PDFs and keep OCR plus a table model for the scans.
Output to CSV or a spreadsheet with the page number and table number in each row.
Add automatic checks for totals, column counts and number formats.
Show a reviewer the page image beside the extracted table for anything that fails a check.
Test on ten of your real tables, including the worst, before processing a batch.
05
What ANVE does today
ANVE lists vision models that read scanned documents with OCR, and a Bills function that reads GST bills. It does not list a general table-extraction function, and it publishes no table accuracy figures. Both of the listed functions run today in the development build; no sealed unit exists yet.
06
Questions people ask
How do I know whether my PDF is scanned?
Try to select a line of text in a PDF viewer. If you can, the PDF has a text layer. If you cannot, the pages are images and need OCR first.
Can a language model read a table from a PDF directly?
It can if it sees the page as an image, or if the table survives as text. Rule-based libraries use the position of each character to rebuild rows and columns, which a model given only extracted text no longer has.
Are these tools free for commercial use?
Camelot, pdfplumber, Docling and Table Transformer are listed under the MIT licence, and Qwen2.5-VL under Apache 2.0, as of October 2026. Check the licence of any version you install.
What about tables in Word or Excel files?
Those formats keep the cell structure, so a document library can read them without OCR. Docling lists DOCX, PPTX and XLSX among the formats it parses.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
