Library files

Each file can be up to 5 GB, and a Library holds up to 5,000 files or 50 GB in total — the same on every plan, personal or organization. Saving a generated artifact back to the Library is capped at 25 MB.

Executables, installers, and opaque archives (like .exe, .dmg, or .zip) are not accepted — upload the underlying research files directly. PowerPoint files (.pptx) are also not accepted; to build slides, ask for a presentation instead. Multi-file scientific formats (BrainVision .vhdr/.vmrk/.eeg, WFDB .hea/.dat) work best uploaded together into one folder so Cortexa can stage the whole bundle.

Large text and delimited datasets remain stored in full, but the searchable preview is built from the first 4 MB of a text, CSV, TSV, or JSON file, and keeps at most 200,000 characters — otherwise upload processing would stall on a large file. This matters when you ask a question the whole file has to answer, such as counting rows or finding the maximum of a column: ask for the analysis rather than the preview, because analysis tools receive the original file and read it in chunks.

Visual analysis accepts valid PNG, JPEG, WebP, and GIF images up to 5 MB. Cortexa reads an image by looking at it, and does not yet recognize text inside one — a scanned page or a screenshot of a table is described rather than transcribed, so upload the underlying PDF or CSV when the words or figures matter. Cortexa validates the stored image before model analysis and asks for a replacement instead of failing the session when bytes are corrupt, mislabeled, or in an unsupported SVG or unconverted HEIC format.

A file in one of these formats is stored as its original bytes and never decoded as text. Cortexa recognizes it by extension, by content type, or by inspecting the leading bytes, so a recording saved under an unhelpful name is still classified correctly.

Select @ in the composer to open the Library browser. Use Select all on a folder or check individual ready files, then choose Attach files. The message stores a snapshot of the selected file references, so later folder changes do not alter historical chats.

You can also drag files onto the composer, or paste a screenshot straight into it — a pasted image is saved with the date and time you pasted it, so several are still tellable apart. A copy that carries both an image and text (a spreadsheet range, say) pastes as text.

Files you attach with the paperclip in chat are session uploads — they stay with that conversation until you delete the session. Click the file name, before or after you send the question, to preview it without leaving the chat. Images and PDFs show the original; documents and tables show the extracted text. The Library uses the same preview — a filename no longer opens a raw storage link. To reuse a session upload in future chats, abstracts, or decks, click Save to Library on the attachment chip (before or after sending). Scope only changes when you click that button; nothing is silently promoted.

A Hugging Face model run saves its JSON result automatically with the model, task, and run identifier. The result holds the model output and its provenance — for example, a text-classification run saves the predicted labels and scores, and a summarization run saves the generated summary — so you can reuse or cite it later.