D2Snap   |   Demo

Downsampling is a technique to reduce data through local, lossy consolidation, such that encoded features are largely preserved. A digital image can be locally consolidated by averaging an n-tile of pixels – the depicted object would remain recognisable. We transfer this principle to the DOM: it can analogously be locally consolidated by merging n-subtrees of nodes into a single ‘average’ node.

D2Snap is an algorithm to downsample the DOM, designed for use with LLM-based web agents. It consolidates user interface (UI) features. At reduced size, the DOM (in serialised form: HTML) shall facilitate eliciting targeted action suggestions from LLMs to progress a specified task. For instance: "Add a Pizza Margherita to a delivery order".

Downsampling is the inverse of quality, governed by ratios: 0: no downsampling. 1: maximum downsampling, i.e., the DOM in its most compressed form at potential cost of certain UI features. D2Snap provides three different downsampling ratios:

• rE: Element downsampling ratio – folding elements in the DOM into each other.
• rT: Text downsampling ratio – truncating text at the ranked sentence level.
• rA: Attribute downsampling threshold – removing attributes by UI-relevance score. With the default attribute scores, which are distributed over UI-relevance and use-frequency, this threshold latently represents an absolute ratio (idempotent).

Actionable elements and minimal text are preserved throughout.

Whether to use the default sentence ranker applying the TextRank algorithm. If disabled, leading sentences are preserved regardless of centrality (high-level relevance).

Whether to apply text downsampling at once to consecutive sibling text nodes (assuming continuous prose). If enabled, text reductions will be more significant, whilst resulting in merged text nodes in the output DOM.

Whether to translate text formatting elements to Markdown. By design, Markdown is a more space-efficient equivalent of basic HTML text formatting syntax.

Whether to pre-process the input DOM by filtering irrelevant features and normalizing the input so the downsampling algorithm works on (more) idiomatic syntax.

Default normalisation includes:

• Filtering: remove comments.
• Filtering: remove elements that are noise (e.g., LINK or SCRIPT).
• Filtering: remove empty elements (cascading up).
• Filtering: remove images that have neither a src nor an alt attribute (empty images).
• Filtering: remove data URL attribute values.
• Normalisation: convert SVG elements to IMG elements to obtain uniform image elements.
• Normalisation: convert icon-font elements (via class attribute like "fa fa-arrow") to images to obtain uniform image elements.
• Normalisation: promote an expressive text-label attribute value (e.g., from aria-label or title) to text of empty elements.

Whether to assign unique IDs to every element in the input and output DOM (only reflected on output in the demo). The ID is an incremental integer written to the data-uid attribute. If enabled, LLM-elicited actions targeting elements in the output DOM via such ID can be traced back to the (live) input DOM. The output DOM is inflated by the encoded IDs. A non-invasive approach could implement a client-side-only source map.