Skip to content
vavio

Can Hermes read a PDF, spreadsheet or image I send it?

Yes. Hermes Agent reads PDFs that have a text layer, Word (.docx) and Excel (.xlsx) files by converting them to text with its read_file tool. Any chat model can work with that text. Images are different. They need a model that can see, either your chat model or a separate image model Hermes calls for a description. In our test on a Vavio computer, a PDF and an Excel file attached in the Hermes app were read correctly with a text-only model. An image was not read until we switched the chat to a model that accepts images.

This follows Hermes Agent v0.23.3, commit caafe53, checked on 21 November 2026. We read the official documentation and source. We then ran one test in the Hermes app on a Vavio QA computer running Hermes v0.23.1, which is described below.

Which files can Hermes read?

You send How Hermes reads it Needs
PDF with selectable text Converted to Markdown text by read_file Hermes’s optional anydoc converter, installed on first use where installs are allowed
Scanned PDF, photo of a page or fax Not readable as text. Hermes warns that the pages are missing. An image-capable model or an OCR route (see below)
Word .docx, Excel .xlsx Converted by read_file Nothing extra
Older or other Office files: .doc, .xls, .pptx, .odt, .ods, .rtf, .epub Converted by read_file The optional anydoc converter
CSV, text, Markdown, JSON, code Read as plain text Nothing extra
PNG, JPEG, GIF, WebP screenshot or photo Shown to the model as an image, or described to it A model that can see images

read_file refuses documents over 50 MB, but the Hermes app in a browser accepts files up to 16 MiB, so that limit applies first. A converted document is read in pages of text, so a long PDF is not loaded into the conversation all at once. The anydoc converter worked on the Vavio computer we tested.

What happens when I attach a file in the Hermes app?

In the Hermes app, which is the chat Vavio opens for you, click + beside the message box. Choose Files… for documents or Images… for pictures. You can also drag a file onto the chat. Each attachment shows as a chip above the message box. Then ask your question in the same message.

  • A document goes to the computer, not into the model. The file is uploaded into Hermes’s own folder on your Vavio computer. Hermes gets a note with its location and opens it with read_file. A text file small enough to fit is included in the message directly.
  • An image goes to the model. If your chat model accepts images, Hermes sends the picture with your message. If it doesn’t, Hermes needs a separate image model. Depending on the Hermes version and app, it either has that model describe the image first, or gives your chat model the file to examine with its image tool. With no image model available, the image isn’t read.
  • Limits in a browser. One file can be up to 16 MiB. SVG and ICO images can’t be attached from a browser. Use PNG, JPEG, GIF, WebP or BMP.

The model’s provider bills the converted text as input, like anything else in the conversation. A 40-page PDF costs more to discuss than a one-page invoice.

What we saw on a Vavio computer

On 21 November 2026 we attached three small files in the Hermes app on a Vavio QA computer: a one-page PDF invoice, an Excel sheet with two rows and a PNG showing a four-digit code. The chat used OpenRouter. We asked for one value from each file.

File Model z-ai/glm-5.3 (text only) Model deepseek/deepseek-v4.1-flash-1017 (accepts images)
PDF invoice Correct reference, read with read_file Not tested
Excel sheet Correct unit count Not tested
PNG image Not read. Hermes reported it could not see the image and had no image tool, then started a background helper that had still not answered when we stopped the test. Correct code, read directly, in 28 seconds

The text-only runs cost about 2 and 10 US cents on OpenRouter; the first was cut short. The image-capable run cost less than a tenth of a cent. We think much of the higher cost came from the helper Hermes started to work around the missing image tool, but the bill does not separate it.

The source explains why. With a text-only chat model, Hermes looks for an image model on other providers it can reach, but it skips the provider you already use for chat. With OpenRouter as the only provider, it found none, so it also hid its image tool. The message Hermes added about not seeing the image is from the Hermes v0.23.1 app; v0.23.3 words this differently, but the same search applies. We did not read the computer’s logs. We ran this test once and only with these models. It shows what happened on that computer, not a ranking of models.

How do I make Hermes read images?

  1. Pick a chat model that accepts images. In the Hermes app, open the model menu in the message box and choose an image-capable model for that chat, or set one as your default in Hermes settings. Your provider’s model list shows which models accept image input. Hermes checks the model’s listed abilities and sends the picture directly when it can.
  2. Or keep your model and name an image model for descriptions. Set auxiliary.vision in Hermes’s settings to a provider and model that accepts images. Hermes then describes every image with that model, even when your chat model could see it. The description call is billed by the provider you name. If OpenRouter is your only provider and your chat model is text-only, you need this step: Hermes won’t pick an OpenRouter image model by itself, so set auxiliary.vision to OpenRouter and an image model.
  3. Retest with a small image that has an unambiguous answer, such as a screenshot of a code. A correct answer in the reply is the sign it works.

When your chat model sees images directly, Hermes re-sends the most recent few images (three by default) with follow-up questions. It removes its kept copy after about 24 hours. After that, or with a text-only model, a follow-up question gets only Hermes’s earlier description or a placeholder.

What about scanned PDFs?

A scanned page is a picture of text with no text layer, so read_file gets nothing from it. Hermes doesn’t treat that as an empty document. When a whole PDF is scanned, read_file returns a NEEDS OCR warning that names the missing pages. It then suggests turning those pages into images and reading them with the image tool, or using an OCR skill. If a Firecrawl API key is saved in Hermes, it sends the PDF to Firecrawl’s hosted OCR instead, unless file_tools.hosted_ocr is set to false. That uses your Firecrawl account and sends the whole document to Firecrawl.

Two limits apply on a Vavio computer today. Neither program below is in Vavio’s setup, and neither was in the system program folder of the QA computer we checked on 21 November:

  • The image route needs both an image-capable model, as above, and the pdftoppm program to turn pages into images.
  • Hermes’s warning for a PDF that is only partly scanned uses the related pdftotext program, and even then it appears only when at least two pages and a fifth of the document, or ten pages, come out empty. Without it, a partly scanned PDF can convert with sections that have headings and no text, and no warning. Check any section that matters.

If a document is scanned, the simplest route is a text-layer copy: export it again from the original program, or run it through OCR before you send it.

What about spreadsheets?

Hermes reads every visible sheet of an .xlsx file and gives each one a heading. It reads each cell’s stored value, not the formula and not the cell’s display format, so a date can appear as a serial number such as 45987 and 15% as 0.15. A formula with no saved result reads as empty. Hidden sheets are skipped; hidden rows and columns are included. It reads at most 5,000 rows and 256 columns per sheet, and the source shows no note to the model when a sheet is cut short. For a larger sheet, ask Hermes to summarise it with a script, or export only the part you need. A CSV is plain text, so Hermes reads all of it, in pages.

What about Telegram, Discord and other messaging apps?

You can also send documents and photos to Hermes in a messaging app. Hermes saves the file on your Hermes computer and tells the model where it is. Small text files are included in the message directly. Photos follow the same image rule: they are sent to an image-capable model or described first. Telegram bots can download files up to 20 MB. We did not test this on a Vavio computer.

Before you send a file

  • Send what the model can use: a PDF with selectable text, an .xlsx rather than a screenshot of a table, a PNG rather than an SVG.
  • For images, check your model first. If Hermes says it can’t see an image, switch the chat to an image-capable model and send it again.
  • Count the size. A long document is billed as input each time Hermes reads part of it.
  • Remember where it goes. On a Vavio computer, the file is saved on your Vavio computer, and its text or image is sent to the model provider you chose.

Vavio runs Hermes on its own computer for $14.99 a month, with model fees billed separately by your provider. See the plan. If you run Hermes yourself, the same settings apply; follow the official document extraction guide.

Sources

Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.3 documentation and source. The Hermes app steps were tested on a Vavio QA computer running Hermes v0.23.1 on 21 November 2026, with one small PDF, one Excel file and one PNG. Messaging apps, scanned PDFs and large files were not tested.

More from the Vavio blog: Hermes Agent guides