Files and documents
LlamaBoss handles each kind of file differently. Small text goes straight into the message; large files and documents are copied into the chat's workspace and read with tools, so they don't flood the model's context.
Three ways to attach
- Click the paperclip and choose files.
- Drag files onto the window.
- Paste a screenshot, an image, or a large block of text.
Up to 10 attachments can go with one message. Images can be up to 50 MB and text up to 64 MiB.
What LlamaBoss recognizes
| Type | Examples | How it's handled |
|---|---|---|
| Images | PNG, JPEG, GIF, BMP, WebP, screenshots | Sent to the model as images. Needs a vision-capable model. |
| Text and code | TXT, logs, Markdown, JSON, source files | Small files go into the message as text. Large ones are stored in the workspace (see below). |
| Reports, policies, fillable forms | Copied to the workspace and read with the PDF tools. | |
| Word | DOCX, DOCM | Text and structure read with the Word tools. |
| Spreadsheets | XLSX, CSV | Inspected with the workbook tools instead of pasting raw rows. |
| ZIP | Source archives, bundles | Listed without unpacking. Extracting needs a separate step. |
Large files become stored variables
When a text attachment or a tool result is bigger than about 12 KB, LlamaBoss saves it to disk and gives the model a short card instead of the whole thing. The card shows the file's size, line count, structure (headings, code declarations, CSV columns or JSON keys) and a preview of the start and end.
The model then uses grep and read_range to pull out just the lines it needs. This is how a small local model can answer questions about a file many times larger than its context window. The chat shows a summary and file/variable cards where appropriate. Tools also have output caps, so a preview is not a guarantee that every byte or row was read. The model can search and read selected ranges from the stored file. As context fills, older tool output and long arguments may also be stored this way.
In Agent Mode, the model can extract text, list the fields in a fillable form, and fill compatible forms. For example:
Inspect this PDF form, tell me which fields it has, and fill them in with the details I give you.Reading and inspecting run without asking. Filling saves a new PDF and leaves the original untouched. Forms built only with XFA may not work.
Word, Excel and CSV
Word: the model can inspect headings, paragraph and table counts, sections, images and styles, then extract readable text and tables. The fixed helpers read documents; they do not render pages or edit the Word layout.
Excel and CSV: the model can inspect sheets and columns, write summary reports, convert CSV to XLSX, and build new workbooks. For example:
Inspect this workbook, remove duplicate records, add a summary sheet, and save it as a new XLSX without overwriting the original.ZIP archives
Attaching a ZIP never unpacks it. The model can list what's inside first, then extract it with path and size checks that stop an archive from writing outside its target folder or exploding in size. Extracted files go into a per-archive folder under Extracted. Ask a vision-capable model to look at extracted pictures with the image-view tool; a text read cannot show them. Each tool call can attach up to eight images of up to 20 MiB each.
Where files go
Each conversation gets a folder you can open in File Explorer:
%USERPROFILE%\LlamaBoss\Chats\<date>_<title>_<id>\LlamaBoss creates folders only as they are needed: attachments for saved images, artifacts for presented files, Workspace for the working directory, and tool-specific folders such as Documents, Spreadsheets, PDFs, Filled Forms, Word, Scripts, Extracted and ToolOutputs. Stored variables live under Workspace\Vars. Older chats may retain their legacy Workflows location if migration could not move them. Files the model creates appear in the chat as file cards and stay on disk after the chat ends.