SideNote Pro

Attach PDFs and Word documents to an AI conversation

Drop a specification, a report or a contract onto SideNote Pro and ask about it. Extraction happens on your machine, the original file is never uploaded, and page markers survive so a PDF answer can point at a page.

Supported formats and limits

Document input at a glance
PropertyValue
FormatsPDF, DOCX, TXT, Markdown (.md, .markdown)
Documents per requestUp to 5
Maximum input file size20 MB
Maximum extracted text2,000,000 characters
Attachment methodsFile picker, drag and drop
Oversized contextRejected with a clear message, never silently truncated

That last row matters more than it looks. Silent truncation is the failure mode where a model answers confidently about a document it only saw two thirds of. SideNote Pro refuses the request instead, so you can shorten the input or raise the context budget knowingly.

Extraction happens locally

Attached files are parsed on your machine and shown as separate attachment chips before you send anything. The original file is never uploaded: what enters the request is extracted text, delimited as untrusted reference material inside the ordinary Chat Completions call.

  • PDF. Text extraction preserves page markers, so a response can attribute a point to a page.
  • DOCX. Document text, lists and tables are read without running macros, loading external resources or pulling in embedded objects.
  • TXT and Markdown. Read directly.

There is no OCR. A scanned or image-only PDF contains pictures of text, not text, and is rejected with a clear message rather than sent as an empty document. If you need to ask about a scanned page, capture or attach it as an image instead and use a model with image support.

The privacy boundary, stated precisely

Local extraction is a genuine privacy property, and it is also easy to overclaim. Here is the exact boundary.

Stays on your machine

The original file. It is read, parsed and closed. SideNote Pro never transmits the file itself, and never opens a network connection during extraction.

Goes to your provider

The extracted text, once you press Send. If the document is confidential, that text reaching your provider is the thing to think about, not the file format.

Extracted text and document metadata are stored inside the conversation archive, so a saved thread still makes sense later even if you moved or deleted the original file. Enabling history protection applies DPAPI to document contents along with everything else. Markdown export lists attachment names without dumping their extracted bodies.

Getting good answers from documents

  • Ask for structure, not just a summary. *Extract every dated commitment and who owns it* beats *summarize this*.
  • Attach two documents to compare them. Two proposals, two contract versions, a specification and its implementation notes. Comparison is where the five-document limit earns its keep.
  • Point at the part you mean. With page markers preserved, *what does the section on data retention actually require* works better than making the model guess which section you care about.
  • Watch the context budget. A large PDF plus a long conversation can exceed it. Managing conversations explains compaction and how to raise the budget.
  • Verify before you rely on it. For anything with legal or financial consequence, an AI reading of a document is a first pass, not an authority.

For a walkthrough with concrete prompts, see asking AI about PDF and DOCX files. If you find yourself attaching the same set of files repeatedly, local folder context is the better tool.

Keep reading