Optimize new snapshot system

[braden-seaborn]

Snapshot Data Optimization & Compression

Objective

Implement configuration-driven filtering and compression so each snapshot keeps only the data an agent truly needs, hitting our low-token budget without losing interactive context.

---

Technical Approach

1. Configuration-Based Exclusion System

• Negative Filtering – declare what to drop instead of what to keep.
• Hierarchical Rules – exclusions can apply at any DOM level.
• Dynamic Filtering – decide at runtime using element properties.
• Token Budget Management – hard limits prevent snapshot bloat.

2. Data-Compression Strategies

Element-Level Compression

{
  "exclusions": {
    "unnecessaryAttributes": ["data-*", "style", "class"],
    "emptyElements": true,
    "hiddenElements": true,
    "decorativeElements": ["img", "svg", "span.icon"]
  },
  "compression": {
    "textTruncation": 100,
    "attributeShortening": true,
    "positionRounding": 2
  }
}

Structural Optimization

• Flattening – drop useless nesting around leaf nodes.
• Deduplication – merge identical siblings.
• Reference Compression – replace verbose IDs with short hashes.
• Selective Depth – stop traversal early in non-interactive branches.

---

Implementation Features

Configuration Schema

• Exclusion patterns for tags, classes, attributes.
• Compression levels: aggressive, balanced, minimal.
• Token-budget ceiling with overflow handling.
• Whitelist / blacklist support for must-keep elements.

Compression Algorithms

• Attribute Filtering – strip non-essential attrs.
• Text Compression – truncate with ellipsis beyond N chars.
• Position Optimization – round coords/dims to N decimals.
• Reference Encoding – generate minimal unique refs.

Smart Exclusion Rules

const exclusionRules = {
  // Ignore low-value elements
  skipIfEmpty: ['div', 'span', 'p'],
  skipIfHidden: true,
  skipDecorative: ['.icon', '.decoration', '.spacer'],

  // Text handling
  textLimit: 100,
  preserveKeywords: ['submit', 'apply', 'next', 'continue']
};

---

Expected Optimizations

Optimization Type	Before	After (Example)	Token Savings
Unnecessary Attributes	100 % captured	≈30 % filtered	~70 %
Decorative Elements	All included	Smart exclusion	40-60 %
Text Content	Full strings	Truncated/summarized	50-80 %
Positional Data	Exact coords	Rounded values	20-30 %

---

Configuration Examples

Aggressive Compression

{
  "mode": "aggressive",
  "tokenBudget": 500,
  "preserveOnly": ["interactive", "form", "navigation"],
  "textLimit": 50
}

Balanced Optimization

{
  "mode": "balanced",
  "tokenBudget": 1000,
  "excludeDecorative": true,
  "textLimit": 100
}

---

Performance Targets

• Token Reduction – 60-80 % smaller snapshots.
• Processing Speed – <100 ms on typical pages.
• Accuracy Retention – 100 % of interactive elements kept.
• Compression Ratio – tunable per project needs.

This optimization pipeline ensures our snapshots stay lean, cost-effective, and fully actionable for the agent.

[braden-seaborn]

• Need to allow interactive excluded tags to be kept
• Need to be able to scan siblings of an object for duplicates

[braden-seaborn]

Should look into vectorizing and fragmenting snapshots that are above a certain threshold to lower token usage.