Supported formats¶
| Format | Extension | Notes |
|---|---|---|
.pdf |
Text layer and, where needed, text inside images | |
| Word | .docx |
Office Open XML only |
| PowerPoint | .pptx |
One page number per slide |
| Markdown | .md, .markdown |
Headings become section citations |
| Plain text | .txt |
What is not supported¶
Legacy Office formats (.doc, .ppt) are rejected. They need a
converter process Primer does not run, and accepting them would mean failing
them after the upload rather than before it.
Spreadsheets, images, and audio are not accepted. A spreadsheet's meaning usually lives in its structure rather than its prose, and answering from it well needs different machinery than answering from a document.
How the format is decided¶
Primer decides what a file is from its contents, then checks that against its name. Both have to agree.
Trusting the extension alone would hand the parser a PDF wearing a .txt
name. Trusting the contents alone would fill a library with files whose
names promise something they are not. A mismatch is rejected with
unsupported_content.
Text inside images¶
Slides and scanned reports carry much of their content as pictures. Primer reads text out of images so that content is searchable.
Recognised text is a transcription
Text read from an image is a machine's reading of it, not the document's
own characters, and it can be subtly wrong in ways that are invisible in
the result. In testing, an image reading SCANNED EVIDENCE was
transcribed as SCANNED E EVIDENCE.
Primer does not currently mark which passages came from images, so a
citation cannot tell you that it was transcribed. If exact wording
matters, check the source. Operators who prefer exactness over coverage
can switch image reading off; scanned documents are then reported as
ocr_required rather than being indexed approximately.
Documents with no readable text¶
A file that parses but yields nothing to index is reported rather than silently accepted:
ocr_required— there is no text layer, and image reading is switched off.no_text_found— images were read too, and there was still nothing.
An empty index for a document would otherwise look like success, and the document would simply never appear in an answer.