Neighbor-aware embeddings for contextual disambiguation

[luigiagent]

Problem

HashingEmbedder embeds each element independently. "Search" next to a text input means something different from "Search" next to a results heading. Two identically-named elements in different page regions produce identical embeddings.

Proposal

Include neighboring elements as weak signals:

func (h *HashingEmbedder) vectorizeWithContext(
    el ElementDescriptor,
    prev, next *ElementDescriptor,
) []float32 {
    vec := h.vectorize(el.Composite())
    if prev != nil {
        nVec := h.vectorize(prev.Composite())
        for i := range vec { vec[i] += nVec[i] * 0.1 }
    }
    if next != nil {
        nVec := h.vectorize(next.Composite())
        for i := range vec { vec[i] += nVec[i] * 0.1 }
    }
    h.normalize(vec)
    return vec
}

Inspired By

• mgrep: context-aware parsing where chunks carry surrounding context
• Word2Vec/CBOW: words defined by their context

Acceptance Criteria

• EmbeddingMatcher.Find() accepts or computes neighbor context
• Neighbor influence configurable (default 10%)
• "Search" near textbox gets different embedding than "Search" near heading
• Nil neighbors = current behavior
• Stays sub-millisecond per element