Image Attachment Extraction

A lot of institutional knowledge arrives as pictures — a photo of a whiteboard after a workshop, a screenshot of a dashboard, an org chart, or a scanned page. Sera does not skip these. It reads image attachments and folds what it sees into memory automatically.

What it does

When an email carries an image attachment in a supported format — JPEG, PNG, GIF, or WebP — Sera:

1. Reads the image using Claude's vision capability.
2. Writes a detailed text description of the image and appends it to the email body as a clearly labeled section, so the visual content becomes text the rest of the pipeline can work with.
3. Runs normal extraction over the combined body, exactly as it would for any email.

The result is that a whiteboard photo, a screenshot, an org chart, or a scanned document becomes searchable memory — the tasks, names, decisions, or structure visible in the image are extracted like any other content.

Predictable cost by design

Vision is handled by the lower-cost Haiku model specifically, rather than the primary extraction model. That keeps the cost of describing an image predictable and modest even when messages carry several attachments. A short timeout guards each vision call so a slow or oversized image cannot stall the poll cycle.

Failures are non-fatal

Image reading never blocks the rest of processing. If Sera cannot make sense of an attachment — it is corrupt, unreadable, or the call times out — it does not discard the email or halt. Instead it appends a note stating the image could not be extracted and that manual review is recommended, then continues extracting everything else in the message. You lose the automatic read of that one image, nothing more, and you get a visible flag telling a human to look.

Two things to know about formats

• Animated GIFs are handled by reading the first frame only. A single representative still is described; the animation itself is not analyzed frame by frame.
• PDFs are not treated as images. They go through a separate text path, not the vision path, because a PDF's text can be read directly and more accurately than a picture of it.

Zero configuration

There is nothing to switch on. Image attachment extraction is active by default for every message that arrives with a supported image. Nobody needs to change how they send things. Attach the photo or screenshot as you normally would, and Sera turns it into memory. The one habit worth keeping is to glance for the "manual review recommended" note when an image is important, so anything the automatic read missed still gets a human pass.

Key points

Sera reads JPEG, PNG, GIF, and WebP attachments with Claude vision, appends a labeled text description to the email body, then runs its normal extraction so images turn into searchable records. It uses the lower-cost Haiku model specifically for vision to keep costs predictable, with a short timeout. Failures are non-fatal: it appends a "could not be extracted - manual review recommended" note and continues. Animated GIFs use the first frame, and PDFs go through a separate text path rather than vision. There is no configuration; the feature is on by default.