// local_documents
Your files stay on the device. The sources have to be real.
Attach one file or several to a chat. A knowledge base is a named library you can reuse. Nothing is uploaded.
- 1 Attach one file or several
- 2 Read and indexed on device
- 3 Answer from matching passages
-
// ingest
Read on your device.
Attach a PDF, a Word document, or a PowerPoint deck. Mortise takes the text out on your machine, and each format contributes what it actually stores. Word gives headings, paragraphs, lists, tables, and image descriptions. PowerPoint is read from the deck file itself: tables, the numbers cached in charts, SmartArt text, picture descriptions, and speaker notes.
Scanned PDF pages — the kind that are a photograph of a page — go through Apple's on-device text recognition, in English. The same recognition runs on a PowerPoint picture when the slide is mostly that image. A Word file is read from its text; pictures inside it are left as descriptions if the file has them. If a file has nothing readable (an image-only Word document, or a picture-only deck with recognition turned off), it simply stays out of search. Recognition is on by default; you can turn it off in Settings → “Read scanned PDFs”.
-
// retrieve
One file can be read whole. Several files are searched.
What Mortise found is split into passages and indexed on the device. Meaning-based search uses on-device embeddings. If those aren't available — missing pieces, or the phone is low on memory — keyword search still runs, so the file stays usable.
With a single attached file, Mortise can use the whole document when it fits, a summary when you're asking about the file as a whole and it's too long, or a search over passages when you need a specific part. With two or more files, or a knowledge base in the chat, it searches: it looks at the words you typed and the meaning of the question, combines those two rankings, and brings in nearby passages so the model has enough context. On a freshly opened knowledge base, the first answer may wait until a search has actually run.
-
// ground
The sources you see come from the search.
In chat, the model is shown the matching passages, labeled with the file name and a page or slide. It's asked to answer from those passages, and to say so when a figure isn't in them. If they only cover part of the question, it can search again. The sources you see are taken directly from that search. Source chips only list files that came back in the results. The wording of the answer is still the model's — including any page number it writes in the text.
If the model claims it read your files, or that you never attached them, and no search actually ran, Mortise replaces that answer with one that did search. You may see the first draft appear, then get swapped out. An answer that never claimed to have looked at your documents is left as it is. Later in a thread that already used your files, Mortise can hold a guess until a search has run.
-
// knowledge_bases
A library you can reuse.
Several files in a chat are enough on their own. A knowledge base is a named collection you keep, so the next conversation can use the same library without attaching everything again. Drop that library into a chat and it uses the same search as files you attached there: keywords and meaning, combined, with nearby passages included.
The search tab inside a knowledge base is a quieter view of the same library. It groups matching passages that sit next to each other and answers from numbered excerpts — [1], [2] — taken from those hits. Attach the library to a chat and you get the fuller chat path, nearby passages and source chips included.
// honest_limits
What you can count on.
Source chips list files that came back from search. Quotes and page numbers in the answer are still the model's words. There are no confidence scores and no separate check that a quote appears in the file.
Text recognition is English. It covers scanned PDF pages and some PowerPoint pictures. Word is read from its text. You can turn recognition off under Settings → “Read scanned PDFs”.
A replaced answer is one that claimed to have read your files when no search ran. You may see that first draft briefly. An answer that never claimed to have looked at your documents stays. Numbered excerpts like [1] are how the knowledge-base search tab presents sources. Chat shows source chips from the search instead.
// sovereign_vault_protocol
Enter the Sovereign Vault
Mortise is coming soon. No account required. Nothing leaves your device.