How to chat with a PDF, Word file or CSV on your Mac, offline

A contract or a sales export does not have to leave your Mac to get questions asked about it. I wanted to know how well the two free models do that, so I tested them. This page has the steps, what I measured, and where it stops working. If you only want the steps, they are the next section.

Attach a file and ask about it

1. Open Outlier and start a chat. Nano and Lite are the free tiers, and they are the two I tested.
2. Click the paperclip beside the message box, or drag the file onto the chat. A chip appears above the box with the file's name, its size and a rough token count. The × on the chip removes it.
3. Type your question and send. If you send with the box empty, the app asks “What's in this” plus the file's name for you.
4. Ask follow-ups in the same chat. The file stays with the chat, up to the size limit under “Where it breaks” below.

The attach button takes PDF, Word (.docx), CSV, TSV, plain text, Markdown, JSON, YAML, TOML and most code files. According to the app's code it also takes Excel (.xlsx) and PowerPoint (.pptx). I did not test those two. You can attach up to 10 files to a message, 50 MB each and 100 MB in all.

What happens to the file

Outlier's window sends the file to the engine on 127.0.0.1, which is your own Mac (what that address is). The engine pulls the text out of the PDF, Word file or spreadsheet and sends it back. The app then puts that text in front of your question and sends the whole message to the model. There is no search index and no summary step. The model reads the entire file every time you ask.

That has two effects. It can quote a figure exactly and add up a column, because the numbers are right there in front of it. And the file has to fit in the model's window, which is why the size limits below matter.

So that follow-up questions can see the file, the app keeps its text in the app's own storage on your Mac, tied to that chat. Delete the chat and the text is deleted with it.

One setting to know about is web search. It is off by default. If you turn it on, your question can reach a search engine, so leave it off when the file is private. Is local AI safe? has more on what stays on the Mac.

What I measured

I wrote five small test files: a 2-page PDF memo, a short Word file of meeting notes, a 12-row sales CSV with a 4-region targets file to go with it, a 5,865-word article and a 26-line Python script with a bug in it. Then I wrote nine questions about them, each with an answer I had worked out before I ran anything. Every question went through the installed app's engine the way the Outlier window sends a turn: the same upload call, the same instructions, the same settings. I ran each question once per model per build, on an M1 Ultra Mac Studio with 64 GB.

A fixed check marks an answer right when it contains the right fact, for example $4,250 or West. For the article summary it wants three bullets with at least two named facts from the article. I read every miss myself.

Table 1: nine questions, build 1.11.919

On 1.11.919 Nano got 8 of 9 and Lite got 9 of 9. The seconds run from sending the question to the last word of the answer.

• Nano's miss is the empty-box send. It answered in one line with no names or dates. I asked that question three more times: two answers were good, and one had the right content but never said it was meeting notes. Nothing Nano said was wrong. It was sometimes thin. It passed once on 1.11.917, missed once on the full 1.11.919 run and passed on 1.11.920, and with one sampled answer each time I can't tell a change from chance. If you want names and dates out of a document, ask for them.
• Every cell is a single answer at temperature 0.7. Run it again and a right answer can turn wrong, or the other way. Treat each cell as one draw, not a rate.

The same nine questions on 1.11.920

1.11.920 is the build the Download button gives you today. I ran the same nine questions on it once: Nano got 9 of 9 and Lite got 9 of 9, including the Nano empty-box send that missed on 1.11.919. The upload code and the window's file handling are identical in 1.11.919 and 1.11.920, so read that as one more draw, not a fix. The three CSV questions took between 5.5 and 21.6 seconds on Nano and between 13.0 and 22.3 on Lite.

The CSV questions, before and after

The three CSV questions are where the free models went wrong at first. On 1.11.917 they slipped on the arithmetic, and a wrong answer came with a visible “let me recalculate”. I scored both builds with the same check. The check takes the first winner an answer names, so a model that said East and then corrected itself to West counts as wrong.

Table 2: the three CSV questions on 1.11.917 and 1.11.919

That is 2 of 6 right on 1.11.917 and 6 of 6 on 1.11.919.

• What changed in 1.11.919: on Nano and Lite, a question that asks for an average, a total, a count or a ranking over the numbers in an attached file now gets the same work-it-out-first step as a typed word problem. The working stays hidden and you see the answer. The math word problems page has the numbers for that step.
• It costs time. Nano's average question took 0.7 seconds on 1.11.917 and 6.6 on 1.11.919. I think six seconds is a fair price for the right number, but it is a price.
• The other six questions barely moved. All six were right on both builds, except the Nano empty-box send above.

Where it breaks

I read these four from the app's code and the messages it shows, and the code is the same in 1.11.919 and 1.11.920. I did not run a file long enough to hit the size limits.

• Scanned PDFs. Outlier takes the text layer out of a PDF. A scan is a picture of a page and has none. The app says it can't read the file and that OCR is not supported yet. Pictures and charts inside a normal PDF are skipped too.
• A big file. For Nano and Lite the app measures a file against a 32,768-token window, at four characters to a token. It warns you past about 24,600 tokens. Past about 30,700 tokens, which is roughly 120,000 characters or 20,000 words, it stops you with three buttons: Remove file, Switch to Core and Send anyway.
• Follow-ups on a bigger file. The app keeps a file's text for follow-up questions only if it is under about 39,000 characters, roughly 7,000 words. Over that it tells you “That file is too large to keep for follow-ups”, and the model sees the file on that one message only. Ask everything in the first message, or attach the file again later.
• A long spreadsheet. CSV and spreadsheet text is cut at 10,000 rows, and the app adds a note when it does. My CSV had 12 rows. I have not tried a big one, so I can't say where the arithmetic starts to slip. If the number matters, check it against the source.

What I did not test

• Core, Quick and Vision 3.8 with files. I tested Nano and Lite only.
• Excel and PowerPoint files. The code accepts them, but I asked no questions about either.
• Anything bigger than my files: a long PDF, a Word file with tables, a CSV with more than 12 rows.
• Other apps. I did not run these nine questions through ChatGPT, Claude or NotebookLM, so I have no comparison to give. For a large library of sources with citations, the NotebookLM comparison says where a cloud tool is the better pick.
• Offline. I ran the questions with the Mac online, so they show what the answers are, not that nothing left the Mac. The check for that is a list of the engine's network connections. The home page shows one taken while a model was answering.

Check it on your own files in ten minutes

You can run my CSV test on any app. Make a CSV of 12 rows and work out three answers yourself: an average, the top item, and which rows miss a target. Then ask the app the same three questions and count the right ones. Do it twice, because a single answer can be luck.

Frequently asked questions