image-vision

Image Vision

Use the doc module with the exact attachment or local path. Do not guess or rewrite the path.

Read One File

Start with the simplest form:

local result = doc.read("/attachments/conversation-id/photo.jpeg")
print(result)

Images and PDFs default to vision when the vision callback is available. The built-in prompt extracts the document as structured Markdown and describes images, charts, diagrams, and other visual elements. This default is appropriate for most requests. Other supported document types default to structural extraction.

Treat the returned text as the visual model's analysis, then answer the user's question rather than merely repeating the extraction.

Supported visual inputs include PDF, PNG, JPEG, WebP, and GIF files.

Read Multiple Files

Issue all asynchronous reads before awaiting any result so vision requests can run in parallel:

local front = doc.readAsync("/attachments/conversation-id/front.jpeg")
local back = doc.readAsync("/attachments/conversation-id/back.jpeg")

local frontText = front:await()
local backText = back:await()
print(frontText)
print(backText)

Customize The Prompt

Set query only when the user needs a narrower analysis, special output format, or more detail than the default extraction:

local result = doc.read("/attachments/conversation-id/chart.jpeg", {
  query = "Extract the chart's title, axis labels, legend, and data values as a Markdown table."
})
print(result)

Preserve the user's request in the query. Set mode = "vision" when an explicit visual-mode override is useful:

local result = doc.read("/attachments/conversation-id/report.pdf", {
  mode = "vision",
  query = "Describe the page layout and all diagrams."
})
print(result)

Use mode = "structural" for a PDF when machine-readable text matters and visual interpretation is unnecessary. When both appearance and exact embedded text matter, run vision and structural reads separately and distinguish their results. Do not substitute structural extraction for requested visual analysis.