Fictura
Hawkings

Attaching files

Give Hawkings an image, PDF, CSV, or Markdown file to read alongside your live data — what it accepts, what it does with it, and what it deliberately never keeps.

The + button beside the Hawkings input attaches a file to the conversation. Use it when the answer depends on something Fictura does not measure — a media plan, a competitor's pricing sheet, an export from another tool, last quarter's board deck.

An attached file belongs to the whole thread, not to one message. Upload a CSV once and every following question can draw on it. Files show as chips above the input; × removes one. ⋮ → New chat starts a thread with none.

What it accepts

FormatExtensionsWhat is read
PDF.pdfThe text layer
CSV / TSV.csv, .tsvEvery row, header included
Markdown.md, .markdownThe full document
Plain text.txt, .textThe full document
Image.png, .jpg, .jpeg, .gif, .webpThe picture itself, up to 8000 pixels on the long edge

Up to 8 MB per file (3.75 MB for an image), 300 pages for a PDF, and 8 files per conversation. Anything else is refused with a reason rather than partially read.

Scanned PDFs will not work

Hawkings reads a PDF's text layer. A scan or a photographed page has no text layer — only an image — and there is no OCR step, so it is rejected as having no selectable text rather than silently returning nothing. If you can't select the words in your PDF viewer, Hawkings can't read them either. Export the source document to PDF again, paste the numbers in directly, or attach the page as an image.

A very large file is truncated to fit rather than rejected, and the chip is marked partial so you know you are asking about part of it. CSVs lose whole rows from the end and keep the header, so what remains is still a valid table.

What happens to the file

A document is never stored as a file. On upload it is parsed in memory, the text is extracted, and the original bytes are discarded before the request finishes. What is kept is the extracted text and a little metadata — the name, the size, and a checksum used to notice you have sent the same file twice.

That is a deliberate design rather than a detail. Nothing you upload is written to disk, saved to a storage bucket, or served back to any browser, so an uploaded file cannot later be re-downloaded, re-opened, or executed by anything.

Images are the one exception. A picture has no text to extract, so its bytes are kept with the conversation. They are never written to a file and nothing serves them back. They leave only as part of a request to Anthropic, so the model can see them.

Two more consequences worth knowing:

  • The type is decided by the bytes, not the name. Renaming an installer to .csv does not get it past the check — the content is inspected, and a file whose contents disagree with its extension is refused.
  • File contents are treated as data, never as instructions. Text inside an uploaded document or image cannot redirect Hawkings, change its rules, or make it reveal anything. If a document appears to be attempting that, Hawkings will say so, which is usually worth knowing about the document.

Attachments are not a data pipeline

Uploading a CSV of revenue does not import that revenue into Fictura. It is context for the conversation and nothing more — it does not appear in your funnel, your ROAS, or any report. To get data into the product, send events or connect the source it comes from.

When a file disagrees with your data

Your live numbers win. If an attached spreadsheet says one thing and the funnel says another, Hawkings treats the tools as the truth about your app — and tells you the two disagree, which is usually the more useful finding.

On this page