ChatGPT and PDF: analyse a file without losing the evidence
ChatGPT and PDF: qualify files, cite pages and check omissions and tables before any decision on a project file.
By Pablo Nastar

A project PDF can mix native text, scans, tables, drawings and appendices. A fluent answer does not prove that the tool read everything. The search for ChatGPT PDF reflects a simple need, getting AI to read a document; but to analyse a file seriously, you first have to qualify the files, split the corpus and require a verifiable reference for every conclusion.
This method suits an initial assisted reading. It does not turn the assistant into an expert on the file and does not replace checking the original document, especially when the decision involves a contract, a drawing or sensitive data.
Contents
- Diagnose the PDF before asking a question
- Build a reading inventory
- Query the file in controlled batches
- Check citations, tables and omissions
- Train the team with a reproducible protocol
- Frequently asked questions
Diagnose the PDF before asking a question
The first check concerns the nature of the file, not the choice of model. A PDF can contain a usable text layer, a scanned image, differently oriented pages or a poorly encoded font. Two pages that look alike on screen can produce very different extractions.
Test three types of page
Choose a text page, a table page and a graphic page. Try to select the text, check the reading order and compare a few words with the original. On a scan, look at the accents, the negations, the units and the numbers. An error on « not », « except » or a dimension can reverse the meaning of a requirement.
The NIST report on text recognition evaluation measures shows why OCR evaluation depends on the document and the chosen measure. It does not provide a universal threshold for your files. The team must identify the errors that really change a business decision.
Classify the at-risk pages
Assign a reinforced control level to drawings, signatures, complex tables, handwritten notes, formulas, dimensions and heavily degraded pages. An unreadable page is not « probably empty ». It becomes an exception to be handled separately.
For teams that first want to understand general formats and uses, the Educasium PDF guide on artificial intelligence provides an entry point. This protocol aims at the traceability of a professional file.
Build a reading inventory
Before any synthesis, create a map of the corpus with the file, the number of pages, the content type and the reading status. This map makes forgotten appendices and excluded pages visible.
| Element | Value to record | Why |
|---|---|---|
| File identifier | controlled name and version | avoid duplicates and old versions |
| Range | first and last page | check coverage |
| Type | text, scan, table, drawing, mixed | choose the processing method |
| Sensitivity | public, internal, personal, contractual | decide whether upload is authorised |
| Status | read, partial, unreadable, excluded | do not confuse absence with non-reading |
| Validator | person responsible for the check | assign the decision |
Normalise without destroying the source
You can split a file into batches, straighten pages or apply OCR. Always keep the original intact and a link between the extracted text and the source page. Do not overwrite the received file with a transformed version that has no history.
The France Num-CNIL sheet on generative AI solutions asks companies to pay attention to the data entered into the tool and to check the outputs. A PDF can contain personal, commercial or financial data that the file name does not reveal.
Query the file in controlled batches
One question per batch and an imposed response format are worth more than a global request for a summary. The aim is to know where the answer comes from and what remains out of scope.
Start with the index, not the conclusion
First ask for the list of detected topics, with the corresponding pages and the unreadable zones. Check that index against a sample. If the references are wrong, the synthesis cannot be considered reliable.
Only then ask a bounded question: which obligations concern a given role, which dates appear in a given appendix, or which documents mention a specific condition. Ask for a column output: statement, short citation, file, page, level of uncertainty.
Work in six passes
Step 1: inventory. Draw up the list of files, versions and pages.
Step 2: qualify. Identify scans, tables, drawings and sensitive data.
Step 3: extract a pilot batch. Choose a small representative set and test the fidelity.
Step 4: query. Limit the question and require page references.
Step 5: verify. Open each cited page and look for omissions in the vicinity of the passage.
Step 6: record. Save the corrections, exceptions and uncovered pages before moving to the next batch.
Hypothetical implementation. A team could create a dummy file with native text, a poor-quality scan and a table spread over two pages. The participants would have to produce a coverage index, refuse a conclusion drawn from the unreadable page and cite the original. This is an exercise, not an Educasium client deployment.
Check citations, tables and omissions
A correct citation can support a local statement; it does not prove that the file contains no exception elsewhere. Verification must therefore address both what is cited and what might have been omitted.
Check the context around the page
Read the previous paragraph, the next one and the cross-references. A condition can be found in a note or an appendix. For a table, check the header, the unit, the row and the continuity on the next page. For a drawing, go back to the visual: a textual transcription does not always preserve the spatial relationship.
Look for false silences
When the tool answers « no mention », ask which pages were actually read. Run an independent search in the file and inspect the lexical variants. An absence is only proven within a defined and correctly processed scope.
The NIST profile for managing generative AI risks proposes measurement and management actions adapted to the context. It does not certify a particular tool and does not exempt you from a test on your own documents.
Train the team with a reproducible protocol
The expected skill is not to obtain a nice synthesis: it is to spot the unread page and prevent an unsourced conclusion from circulating. The training must give practice in refusal and escalation.
Distribute the roles
One participant prepares the corpus, another queries, a third checks the pages and a fourth plays the decision-maker. Swap the roles in the second exercise. This rotation shows that the same answer is not judged the same depending on the responsibility carried.
Explicit absence of client experience. Educasium does not claim here an accuracy rate obtained on client files. Performance depends on the file, the OCR, the service and the question. The method aims to make the limits visible before an answer is used to decide.
The guide on the confidentiality of company data completes this protocol for choosing admissible files. A successful exercise is measured by the errors detected, the sources correctly recovered and the decisions suspended in time, not by a promise of gain.
Limits and counter-cases
Do not use an external service when the contract, professional secrecy or data sensitivity forbids it. Also avoid global analysis when the file contains drawings or tables that the tool cannot faithfully reproduce. In those cases, human reading or a controlled environment remains necessary.
Another limit relates to the cost of verification. If every statement requires a longer review than direct reading, the tool brings no value for that file. This finding is not a failure: it avoids industrialising an unsuitable method.
Frequently asked questions
Which tool should I choose to analyse a PDF?
Choose based on the file type, the sensitivity, the citation capabilities and the service contract. Test several representative pages before generalising. Check the export of results and the possibility of deleting the data. No tool name replaces this test.
Is a page reference enough?
No. It lets you find the passage, but you must check the context, the appendices and the cross-references. A page can also be wrongly numbered after assembly. Keep the file identifier and, if needed, the internal number and the displayed number. The decision remains human.
How should a scanned drawing be handled?
Treat OCR as an aid for locating, not as a reliable geometric reading. Check dimensions, legends and spatial relationships on the original image. Flag blurry zones and invisible layers. Critical data must be reread by the competent person.
Can a client file be uploaded?
Only if the organisation has verified the authorisation, the service contract, the data present and the applicable settings. Minimisation or anonymisation can reduce the risk without always removing it. If the decision is unclear, do not upload. Pass the point to the competent manager.
Make every conclusion verifiable
To analyse a PDF file with AI, keep three traces: what was provided, what was read and the page that supports the conclusion. Everything else remains a hypothesis to be checked.
The introduction to AI and prompting training can include a document-analysis workshop on an authorised corpus. Training eligible for OPCO/FIF PL funding subject to conditions. Qualiopi-certified organisation. To define the programme and the prerequisites, contact Educasium.