Can AI understand arrows in diagrams?

## What AI understands about arrows

In the context of a whiteboard diagram, an arrow typically conveys one of several relationships:

• Flow — data, control, or process moves from A to B
• Causality — A causes or leads to B
• Dependency — B depends on A
• Sequence — A comes before B
• Definition — A is described by B (annotation arrow)

Modern large multimodal models — GPT-4o, Claude 3.5, and the model powering BoardSnap — recognize arrows as directional relationships and attempt to describe the relationship based on the label, the shapes connected, and the context.

## When arrow interpretation works well

• Single-direction arrows between clearly labeled boxes — e.g., [User] → [API] → [Database]. AI reads this reliably and describes the flow.
• Annotated arrows with clear labels — "HTTP POST", "async", "reads only" — labels on arrows are read and incorporated into the description.
• Standard diagram conventions — diamond shapes for decisions, rectangles for processes, cylinders for databases. When you use standard symbols, AI applies the correct semantic interpretation.

## When arrow interpretation fails

• Bidirectional arrows (⟷) — often misread as two separate arrows or one directional arrow. Context helps, but it's not reliable.
• Crossing arrows — the spatial intersection confuses positional tracking. AI may attribute an arrow to the wrong source or target.
• Unlabeled arrows in dense diagrams — without labels, the relationship has to be inferred entirely from box context, which degrades accuracy for complex diagrams.
• Small arrowheads — at low image resolution or on an angled photo, small arrowheads may not be detectable. Perspective correction (via VisionKit in BoardSnap) helps significantly here.

## BoardSnap's handling of arrows

BoardSnap AI reads diagrams in the context of whiteboard content — it expects boxes, arrows, and labels. When it sees an arrow between [Frontend] and [Auth Service], it describes the relationship as data or control flow between those components. Open questions or TBD arrows are flagged as action items.

The perspective correction step is important for arrows specifically: on an angled photo, arrowheads at the edges of the board are more distorted than center content. A flat, corrected image from VisionKit gives the model cleaner input for arrow detection.