Understanding an image
Connect reads PNG, JPEG and WebP. An image is sent to the model as an image, alongside your question and whatever record you were looking at, and the answer comes back as ordinary text you can check against the picture. There is no separate OCR service, no face matching and no image search: what you get is one model's reading of one picture, cited to the file it came from.
What the model is handed#
An image is not converted to text first and then read. It goes to the model as an image, which is why a photograph of a delivery note and a screenshot of a bank statement both work, and why neither produces a tidy machine-readable extract unless you ask for one in the question.
| Sent | Not sent | Why |
|---|---|---|
| The image itself | A transcript of it made beforehand | There is no separate text step to go wrong |
| Your question, as you typed it | A rewritten or expanded version of it | A question about the top-left corner has to stay a question about the top-left corner |
| The record you had open, if the Assistant was opened on one | Every other record in the workspace | Screen context is checked against the database before it is trusted |
| Nothing from another workspace, ever | — | media_assets is scoped and under row-level security like every other table |
The practical consequence is that an image answers a question about itself. Asking whether the invoice in a photo matches the deal on the pipeline is two questions, and Connect answers the second one by looking the deal up with a tool, not by reading it off the picture.
What a picture cannot tell you#
- Who a person is. Connect does not match a face against anything. A photograph of a person is described, not identified.
- Where a photo was taken, beyond what is legible in it. Connect does not read location metadata back to you as fact.
- Whether the image is genuine. A model reads what is in front of it; a doctored screenshot reads as a screenshot.
- Numbers you need to be exact. A model reading a column of figures off a photograph is a reading, not an extraction. If the numbers matter, send the XLSX or the CSV and let table extraction do it.
Where an image is the right thing to send#
- A photograph of something physical
- A damaged part, a meter reading, a handwritten note at a counter. There is no better source than the picture.
- A screenshot of an error
- The exact wording of a message, in context, which a person retyping it usually loses.
- A scanned page with no digital original
- Reasonable, but a scan saved as PDF is better: reading text out of a document tries the text layer first and only falls back to reading the page as an image.
- A logo or a layout you want described
- Fine. Connect describes what it sees; it does not tell you whether the design is any good.
When the reading is wrong#
The failure is almost never dramatic. A model reading a poor image produces a confident sentence that is subtly wrong — a transposed pair of digits, a date read from the wrong line, a total that belongs to the row above. It does not look like an error, which is why the answer names the file it read.
Open the file in preview beside the answer.
Result You are comparing the claim against the source, rather than against your memory of the source.
Ask the narrower question — one field, not the whole document.
Result A single value is far easier to verify than a summary, and a model asked for one thing tends to look in one place.
If the image is genuinely unreadable, say so and attach the original file instead.
Result Connect answers from the document rather than from a picture of it, and the answer stops depending on image quality.
Connect saying it cannot make out a value is a correct answer, not a failure. Treat a reply that guesses at a blurred figure as the worse outcome of the two.
Questions#
Does Connect run OCR on an image?
Not as a separate step. There is no OCR service in the path — the model reads the picture directly. That is why the result is prose you can question rather than a text file you have to trust, and why exact figures are better taken from the original spreadsheet or PDF than from a photograph of one.
Which image formats are accepted?
PNG, JPEG and WebP. Anything else is refused at validation before it is stored, with the reason shown. A file whose extension says PNG but whose content is something else is refused on its content, not on its name.
Can Connect compare two images?
You can attach more than one and ask about them together in a single question. The answer is one reading of both, so it inherits the same limits: good for differences you could point at, unreliable for small numeric detail.