| name | vision-enrich |
| description | Guides autonomous decisions on when extracting and analyzing web images adds sufficient value to justify the token cost for visually dense research domains. |
| version | 1.0.0 |
| tags | ["research","multimodal","vision"] |
Vision Enrich Skill
Overview
This skill provides the decision framework for autonomous visual data enrichment during deep research. It teaches the LLM when to extract and analyze images from web pages, how to evaluate the cost/benefit of visual data, and what constitutes high-value visual evidence.
The skill does NOT mandate image extraction — it provides heuristics so the LLM can make an informed autonomous decision per research context.
When to Use This Skill
Always load this skill when the research topic falls into these domains:
High-Value Visual Domains
| Domain | Why Visuals Matter | Example Triggers |
|---|
| Financial Analysis | Charts contain revenue/earnings data impossible to describe in text | "analyze NVIDIA quarterly earnings", "compare cloud provider growth" |
| Data & Statistics | Distributions, trends, and correlations are best understood visually | "COVID mortality trends", "global temperature change data" |
| Architecture & Systems | System diagrams convey structure more efficiently than paragraphs | "microservices architecture", "neural network architecture" |
| Scientific Research | Figures, microscopy, and experimental results are core evidence | "CRISPR mechanism", "protein folding structure" |
| Product & UI Analysis | Screenshots reveal design patterns and UX decisions | "compare landing page designs", "SaaS dashboard analysis" |
| Geographic & Spatial | Maps and spatial data require visual representation | "supply chain routes", "demographic heat maps" |
| Engineering & CAD | Blueprints and schematics are the primary information carrier | "circuit diagram", "mechanical assembly drawing" |
When NOT to Use Visual Enrichment
- Pure text analysis (literature review, philosophy, legal analysis)
- Simple factual queries ("What is the capital of France?")
- Topics where images are likely decorative (lifestyle blogs, generic news)
- Research with very tight token budgets
Core Principle
Vision tokens are 10-100x more expensive than text tokens. Use them only when images carry information that cannot be adequately conveyed in text.
A financial chart showing revenue growth from $10B to $60B over 5 years is worth the tokens. A stock photo of a smiling executive at a conference is not.
Decision Framework
Step 1: Assess Visual Information Density
Before calling extract_web_images, ask yourself:
1. Does this page likely contain DATA-HEAVY images (charts, diagrams, tables)?
→ Look for signals in: page title, URL patterns (/chart/, /dashboard/, /report/),
search snippet mentions of "chart", "figure", "graph", "table"
2. Would seeing the image change my understanding or just confirm the text?
→ "Revenue grew 45% YoY" — the text tells you the number
→ "Revenue grew 45% YoY but the growth rate is decelerating" — the CHART
shows the trend shape that text cannot fully capture
3. Can I extract the key insight from text alone, or is the visual pattern critical?
→ Text: "The architecture has 5 services"
→ Visual: The ARCHITECTURE DIAGRAM shows data flow direction, bottlenecks
Step 2: Estimate Token Budget
extract_web_images cost model:
- Fetch HTML: ~0 tokens (HTTP, not LLM)
- Download images: ~0 tokens (HTTP, not LLM)
- Encode images: ~1K-150K tokens per image (depends on dimensions)
- LLM vision analysis: ~same as encoding cost
- Total per image: ~2K-300K tokens
Decision threshold:
- Simple topic (2-3 search calls budget): SKIP image extraction
- Standard topic (5-6 search calls budget): Extract 1-2 images IF high visual density
- Deep topic (unlimited): Extract 2-3 images when domain matches high-value list
Step 3: Execute or Skip
EXTRACT when:
- Domain is in the High-Value Visual Domains table above
- Search snippet or URL suggests data-heavy content
- The visual pattern (trend shape, distribution, structure) is the information
- Research depth is "standard" or "deep"
SKIP when:
- Text already captures all relevant information
- Images are likely decorative (stock photos, banners)
- Research depth is "simple" (limited token budget)
- Page is primarily text with no data visualization signals
Step 4: Post-Extraction Analysis
After images are extracted and injected into context:
- Describe what you see: Extract specific numbers, trends, patterns
- Cross-reference with text: Does the visual confirm or contradict the textual claims?
- Cite the source: Include the image URL and page URL in findings
- Note limitations: If the image is partial, low-resolution, or ambiguous, state this
Available Tools
When this skill is active and vision_enrich_data is enabled:
| Tool | Purpose | When to Use |
|---|
extract_web_images | Extract images from web page HTML | After finding a promising URL in search results |
view_image | View a local image file | When user uploads or references an image file |
Workflow Example
Financial Analysis Example
User: "Analyze NVIDIA's Q3 2026 earnings"
Research flow:
1. Search: "NVIDIA Q3 2026 earnings results" → finds results
2. Search result analysis:
- investopedia.com/nvidia-earnings → likely has revenue charts
→ DECISION: EXTRACT images (high visual density domain: financial)
3. extract_web_images("https://investopedia.com/nvidia-earnings", max_images=2)
→ Extracts: revenue bar chart, segment breakdown pie chart
4. LLM sees images + text together
→ Analysis: "Revenue reached $35.1B (chart shows 94% YoY growth,
accelerating from 78% in Q2). Data center segment is 87% of total
(pie chart) — up from 80% last quarter."
5. Final compressed research includes visual + textual analysis
Architecture Research Example
User: "Explain the Transformer architecture"
Research flow:
1. Search: "Transformer architecture diagram" → finds results
2. Search result analysis: arxiv paper page or blog with architecture diagram
→ DECISION: EXTRACT (domain: architecture/system design)
3. extract_web_images(url, max_images=1)
→ Extracts: encoder-decoder architecture diagram
4. LLM sees diagram + text → describes the visual flow
5. Findings include both textual explanation and diagram-based description
Quality Checklist
Before finalizing visual-enriched research:
Common Mistakes to Avoid
- ❌ Extracting images from every search result page indiscriminately
- ❌ Calling extract_web_images for decorative or stock photos
- ❌ Describing images vaguely ("there is a chart showing growth") — extract specific data
- ❌ Ignoring images after extraction — analyze them thoroughly
- ❌ Using visual enrichment for simple factual queries
- ❌ Exceeding reasonable token budget (3+ images when 1-2 would suffice)
Configuration
This skill's tools are controlled by SoulSearcher configuration flags:
SOULSEARCHER_MODE=deepresearch
supports_vision=true
VISION_ENRICH_DATA=true
When VISION_ENRICH_DATA=false, the extract_web_images tool is not available, and this skill functions as a passive guide for manual visual analysis only.