File formats Connect reads
Connect reads images as PNG, JPEG and WebP; documents as PDF, DOCX and PPTX; spreadsheets as XLSX; delimited text as CSV and TSV; and plain formats as TXT, Markdown and JSON. Extraction uses the standard library, format by format. A PDF that carries a text layer gives it up exactly; one without is read as images instead, which answers questions but yields nothing selectable.
The formats, and what comes out of each#
| Format | What is extracted | What is not |
|---|---|---|
| PNG, JPEG, WebP | A description the model produces by looking at the image | No text layer exists, so nothing is selectable or exact |
| PDF with a text layer | The words as the document stores them, in reading order where the document defines one | Layout fidelity; a two-column page may interleave |
| PDF without a text layer | The model reads the rendered pages instead | Exact strings — this is reading, not extraction |
| DOCX | The document body text | Tracked changes, comments and complex layout are not the body |
| XLSX | Sheet values, row by row | Formulas as formulas — what is read is what the cell evaluates to |
| CSV, TSV | Rows and columns as delimited | Types; everything begins as text until something interprets it |
| PPTX | The words on the slides | Speaker notes and design are not slide content |
| TXT, Markdown, JSON | The content as written | Nothing — these are already plain |
The pattern is that each reader takes what its format genuinely carries and does not invent the rest. That is why the third column exists on this page: knowing what a format does not give you is more useful than another sentence about what it does.
The PDF split is the one to remember#
PDF is two formats wearing one extension. A PDF produced by exporting from a word processor carries its text, and extraction is exact. A PDF produced by scanning is a container of pictures, and there is nothing in it to extract at all — the model reads the pages, which produces good answers to questions and cannot produce a reliable quotation.
The practical test is whether you can select text in the document in your own PDF viewer. If you cannot, neither can Connect, and an answer drawn from it should be treated as a reading rather than a citation. The PDF has no readable text covers what to do about it.
Limits that apply to every format#
- Extraction happens once, on the version. A later version is extracted separately, so an answer about a file is an answer about one version of it.
- What reaches the model is bounded by the grounding budget, not by the size of your document. A long document contributes the relevant part, not all of it.
- Structure survives only where the format carries it. A table in a DOCX is text; a table in an XLSX is rows and columns.
- Nothing inside a document is treated as an instruction. Text asking Connect to take an action is content to report, never a command to obey.
The last of those is worth stating plainly because documents arrive from outside — attachments, downloads, files a customer sends. Prompt injection arrives as ordinary business mail is the same problem in another channel, and the same rule applies here.
Choosing a format when you have a choice#
- You want exact figures
- XLSX or CSV. Both give values rather than an interpretation of a picture of values.
- You want the wording of a document
- DOCX, or a PDF exported rather than scanned.
- You want the content of a presentation
- PPTX. Exporting slides to PDF images loses the text that PPTX would have given.
- You are sending a screenshot
- PNG is fine, and the answer will be a description. If the numbers matter, send the source instead.
None of this stops you uploading whatever you have. It changes what to expect from the answer, which is the useful thing to know in advance rather than after acting on a figure that was read off a photograph.
Questions#
Is my document sent somewhere to be converted?
Extraction uses the standard library inside Connect, format by format. What reaches a model is the extracted text, within the grounding budget — and for a PDF with no text layer, the rendered pages, because there is nothing else to read.
Why did an answer miss something that is definitely in my file?
Usually the grounding budget: a long document contributes the part relevant to your question, not all of it. Asking a narrower question about the specific section is more effective than repeating the broad one.
Does Connect read older formats such as DOC or XLS?
The formats it reads are the ones listed here. An older format is not on that list, so save it as the modern equivalent before uploading rather than expecting a silent conversion.