Your files in the AI's memory — what OrbitPilot can read and how to ask
What the AI can read
Everything you place in your OrbitPilot Files folder in the File Library (including subfolders you create inside it) becomes part of your AI's memory at the next memory refresh — the content, not just the file name:
- Spreadsheets (.xlsx, .xls, .csv, .tsv) — every cell, with sheet names and row numbers preserved, so the AI can tell you "Sheet Customers, row 12".
- PDF documents — the text of each page, page by page, so answers can say which page a fact came from. (A scanned, image-only PDF has no text to read.)
- Word documents (.docx) — paragraphs and tables, including price lists and schedules inside tables.
- Plain text, Markdown, JSON, RTF, email exports, PowerPoint — read as text, bounded per file.
- Images — listed by name only for now; their visual content is not read into memory yet.
Files anywhere else in your library are never read — the folder is the consent switch, and you move files in or out of it to control exactly what the AI sees. Files attached to a Blink follow that Blink's own privacy switch instead.
How to ask about your files
Ask in your own words, the way you would ask a person:
- "Find the receipt for the deck project I paid Home Depot in September."
- "What is Juliet's phone number? It's in my customers spreadsheet."
- "What does my contract say about cancellation?"
Run a memory refresh (or just log out — memory refreshes automatically) after adding new files, so they enter memory before you ask.
Every file answer carries a link and a warning
When OrbitPilot answers from one of your files, the answer names the file with a working link, says where inside the file the information lives (page, sheet and row, or section) as precisely as reading allowed, and ends with a standing reminder: "I can make mistakes — please open the file and double-check anything important." The AI may interpret and give an opinion about what a document means, but the link and the reminder are always there so you can verify the original yourself.
For developers
Extraction: orbitpilot/pipeline/file_text.py (registry-driven,
bounded reads, content-hash cached — kind doctext2 since the
page/table/row markers). The consent seam is
filemgmt.models.orbitpilot_visible_files_qs(user) — the one gate.
The build enters through filemgmt.files
(default_visible=True since WO-32 Phase A; admins re-gate per plan
on the Source/Plan Assignment page). Citations:
pipeline.retrieve.add_file_answer_support decorates file chunks
with their URL and arms FILE_ANSWER_NOTE;
services.file_citation_note appends the deterministic footer.
Acceptance tests: orbitpilot/test_files_into_memory.py.