Upload files and Microsoft Office documents

Direct attachment or Workspace library?

Use a direct chat attachment for a one-off summary, a quick comparison, or a question about a small set of files. Direct attachments have tighter per-file and per-turn limits because their extracted text travels with the conversation.

Use a Workspace for larger or repeatedly used material. The file is processed once, marked Ready, and then made available to relevant chats inside the same Workspace. This is usually the better choice for handbooks, contract sets, research libraries, and recurring reports.

Ask clearly about spreadsheets

1. 1Name the workbook and sheet or table you want analyzed.
2. 2State whether the first row contains headers and explain any domain-specific abbreviations.
3. 3Specify the calculation, grouping, date range, or output columns you need.
4. 4Ask LLMBase to state missing values, duplicates, or assumptions before drawing conclusions.
5. 5Request a CSV or XLSX Artifact when you need a reusable result rather than a table only in the answer.

Create a new Office file

Ask Chat for an XLSX, DOCX, or PPTX Artifact when the deliverable should be a new Office file. State the audience and structure: sheet names and formulas for Excel, sections for Word, or slide count and speaker notes for PowerPoint.

Creation never edits an uploaded original. Generated Office files can be downloaded or saved as a new Workspace file; in-app Office editing is not included yet.

Edit exact text in an attached PDF

Attach the PDF and name the exact old text and replacement. Chat creates a new downloadable PDF, rewrites the original text operator, and uses the same embedded or PDF-standard font resource instead of overlaying a similar font. The uploaded original stays unchanged.

For a reliable exact-font edit, the text must be selectable and unambiguous, and its font must contain every replacement character. The replacement must fit the original span. Scans, outlined text, unavailable non-standard fonts, signed or encrypted PDFs, missing glyphs, and unsupported layouts are rejected rather than approximated.

Scans, complex layouts, and formulas

PDFs with selectable text generally produce the cleanest results. Scan-heavy PDFs and images may require OCR, and handwriting, rotated pages, or complex multi-column layouts can reduce accuracy. Review important names, totals, dates, and legal wording against the original.

Spreadsheet extraction and previews use cached workbook values and structure; they do not calculate formulas. A newly generated XLSX Artifact can include an explicitly requested formula and asks Excel to recalculate it when opened, but LLMBase does not execute workbook formulas. External data connections, embedded applications, and interactive Excel behavior are not executed.

DOCX and PPTX previews show extracted text rather than the original layout, formatting, or embedded media. Download the original file when you need to verify presentation details.