Can Local AI Understand PDFs, Images, and Spreadsheets?
· By Michel Valenzuela, Founder of Daryl
Yes, some local AI applications can work with documents, images, and spreadsheet data. Support depends on both the model and the software around it.
A model’s capabilities do not automatically become features in every application that uses it.
The application needs to open the file, extract or interpret its contents, and provide the relevant information to a compatible model.
Before relying on an answer, check what the assistant actually received.
Why does file support depend on the application?
A file is a container for information.
A PDF may contain selectable text, scanned pages, tables, and images. A spreadsheet may contain formulas, multiple sheets, and formatting that explains how to interpret the figures.
The application decides how to handle those elements.
It might extract text, render pages as images, parse rows, or use another process.
If important information is lost during that step, the model may answer from an incomplete representation of the file.
Successful upload does not establish successful understanding.
Can local AI read a PDF?
Some applications can extract and analyze PDF content locally.
The result depends on the document.
A straightforward text document is different from a scan or a page with several columns and complex tables.
| PDF content | What to check |
|---|---|
| Selectable text | Whether reading order is preserved |
| Scanned pages | Whether text recognition is available and accurate |
| Tables | Whether values remain associated with the correct headings |
| Charts | Whether the system can interpret the visual content |
| Footnotes | Whether qualifications are included |
| Long documents | Whether the answer uses the relevant sections |
Ask for page references or supporting passages where the application provides them.
Then inspect the details that matter to your question.
Is a scanned document different from a text document?
Yes.
A scan can be an image of a page rather than machine-readable text.
The application may need optical character recognition, often called OCR, or a model capable of interpreting images.
Recognition can make mistakes, particularly with small text, handwriting, unusual layouts, or poor image quality.
Dates and figures deserve close review.
If an answer depends on a scanned amount, ask the assistant to show what it read before using that amount in a calculation or comparison.
Can a local model understand images?
Some models accept visual input. Others accept text only.
Even with a vision-capable model, the application needs to support sending the image in the required form.
A text-only workflow might instead receive an OCR transcript. That can support questions about visible words while omitting layout, color, or other visual relationships.
Check which kind of input the assistant is using.
“What text is in this image?” and “What does this diagram show?” may require different capabilities.
What about CSV files?
CSV files store tabular data as text, making them relatively straightforward to parse.
But interpretation still requires care.
A column labeled “Revenue” does not establish the currency, reporting period, or whether the figures include tax. A blank cell may mean zero, unknown, or not applicable.
Provide definitions when they affect the question.
For example:
“Each row is one completed project. Amounts are in yen. Blank delivery dates mean the project is still open.”
That context can matter as much as the rows themselves.
Are spreadsheets more complicated than CSV files?
They can be.
A workbook may contain multiple sheets, formulas, hidden rows, merged cells, and references between tabs.
An application that extracts displayed values may not preserve formulas. One that reads a single sheet may miss information elsewhere.
Before asking for conclusions, establish the scope:
“List the sheets you can access and identify the columns available for analysis.”
Then check whether the assistant is working with the material you intended.
A confident answer about “the workbook” is unhelpful if only one tab was read.
Can you trust AI to calculate totals?
Use a calculation method you can verify.
A language model generating an answer is different from software executing a formula or running a data-processing tool.
An application may provide those tools, but you should know when it uses them.
For important figures, ask for the calculation method and check the result against the original data.
Also confirm the selection: an accurate sum of the wrong rows is still the wrong answer.
AI can help explain patterns while a spreadsheet or calculation tool handles the arithmetic.
What is a useful first test?
Choose a small file whose contents you know.
Ask the assistant to describe its structure before analyzing it.
For a document:
“Identify the main sections and quote the passage that states the deadline.”
For a table:
“List the column headings and explain how you interpret blank cells.”
For an image:
“Describe what you can read clearly and identify anything uncertain.”
These requests reveal whether the input was interpreted correctly before you build a larger task on it.
Does local file analysis keep the file on your computer?
It can, if all relevant processing happens locally.
But file selection and local model execution do not establish the behavior of every feature.
Extraction, OCR, image interpretation, and connected tools may use different services.
Check where each stage occurs and whether the application saves copies, extracted text, or indexes.
The privacy question applies to the whole workflow.
How does this relate to Daryl?
Daryl is being developed as a personal AI for Mac, focused on local core conversations, voice and typing, and decision assistance.
Working with material people choose to provide is a natural direction for a useful personal assistant.
However, specific PDF, image, CSV, and spreadsheet import capabilities need to be implemented and tested individually. We are not presenting them as confirmed launch features.
Daryl is still in development. Its supported file types and limitations will need to be documented clearly.
A model’s advertised capabilities are the starting point. The application has to deliver a working experience.
Frequently asked questions
Can every local model read images?
No. The model and application both need suitable support.
Does attaching a PDF mean every page was processed?
Not necessarily. Check the application’s limits and ask for references supporting the answer.
Can a CSV replace a workbook?
It can preserve a table’s values, but generally does not preserve workbook formulas, formatting, or multiple sheets.
What should I check first?
Whether the assistant correctly identified the content and structure relevant to your question.
Verify the input before the conclusion
Start by checking what the application can read.
Then inspect the passages, rows, or visual details supporting its answer.
That makes file-based assistance more useful and gives you a clearer basis for deciding what to do next.
Meet Daryl and get the launch announcement