You can hand the Rixoffice agents reference material directly in chat. Attach a file to the composer and the agent reads it as part of the task — summarizing a PDF, pulling numbers from a workbook, matching a design from a screenshot. No copy-paste relay, no second window, no reformatting between apps. The attachment becomes context the agent works with, in the same request that edits your open file.
What you can attach
The attachment reader understands the formats that actually circulate in office work:
| Type | What the agent can do with it |
|---|---|
| Read pages, search text, summarize, extract structure | |
| DOCX | Read blocks, headings, tables and formatting |
| PPTX | Read slide content and structure |
| XLSX | Read ranges, cells and sheet features |
| Plain text | Read and reason over the full text |
| Images | Multimodal understanding of the visual content |
The image support is the notable one: attachments are not reduced to text. Attach a screenshot of a dashboard and the agent can describe the layout, transcribe chart labels or use it as a style reference for a deck. That multimodal reading is what turns a picture from an unopenable blob into usable context, and it changes which files are worth attaching — a photo of a handwritten note is now as legible to the agent as a typed page.
For the document formats, the reader preserves structure rather than flattening everything to a string of text. A DOCX attachment arrives with its headings and tables intact, an XLSX with its ranges and sheet features, a PPTX with its slide order. Structure is what lets the agent answer questions like "what does the third table say" instead of guessing.
How the reading works
Attachments flow into the agent loop through the same skill system as your open file. The docx skill, for example, includes a read_attachment tool, so the agent can open what you attached, examine it and then act — all inside one request.
A typical exchange looks like this:
- You attach a supplier PDF and ask: list every delivery deadline mentioned and add them as a table.
- The agent reads the attachment, searching for dates and commitments.
- It proposes a new table block in your open document, built from what it found.
- You accept the edit, and only that table is inserted — the rest of your document is untouched.
The same pattern works in every app: in RixSheets the agent proposes operations from an attached workbook, in RixSlides it builds slides from an attached outline, in the PDF editor it cross-references an attached report against the pages you are reviewing.
Everyday uses
- Research intake: attach three PDFs and ask for a combined summary of their findings.
- Data merge: attach an XLSX workbook and have its figures become a chart or table in your document.
- Design reference: attach a screenshot and ask for a slide that follows the same visual language.
- Meeting capture: attach a photo of a whiteboard and ask for a cleaned-up action list.
- Cross-format work: attach a PPTX while writing the matching report in RixDocs, so both stay consistent.
Each of these used to be a multi-window workflow. With attachments it is one request, and the output lands as a normal block-level edit you can accept, refine or discard.
Attachments also compose with the agents' strengths. Oracle can read an attached contract and answer questions about it without touching your draft; Prometheus can build a plan from an attached brief; Momus can check your rewrite against the attached source material. The attachment is just context — every agent in the roster can use it in its own way.
Privacy of attachments
Attachments are processed locally. The agent reads them on your machine, and if you use cloud AI, only the text you ask about is sent onward — not your whole file and not the attachment itself. With Ollama selected, the entire exchange stays on your device, including the attachment reading. The full privacy model is described in AI privacy.
That local processing is the quiet difference between Rixoffice and web-based assistants: your supplier contracts and budget workbooks are read where they already live, on your disk. A web chat tool needs the file uploaded to somebody's server before it can read page one; here the first page is read before anything leaves the machine, and the upload — if any — is only the extracted slice relevant to your question.
Tips for better results
- Say what you want from the attachment, not just that it is there. Compare the attached budget with section two of this document produces far better output than look at this file.
- For long PDFs, ask targeted questions rather than one giant summary; the agent searches pages either way.
- Combine attachments with web search when useful: attach last year's report and ask the agent to find this year's market figures to build a comparison.
- Attach early, then iterate — follow-up questions reuse the same attachment without re-attaching.
What to explore next
- AI web search — pair your own files with live research.
- Agent skills — the read_attachment tool in context.
- PDF editor AI assistant — working with PDFs directly as your open file.