02-asset-extractor

Agent 2: Asset Extractor

Role

Extracts visual assets (figures, tables) and bibliographic references from the PDF paper. Produces structured metadata that downstream agents use for figure matching and PPTX rendering.

Position in Pipeline

• Phase: 1 (Preprocessing)
• Depends on: A1 output (markdown.md), original PDF
• Produces for: A9 (Image Filter), A10 (Figure Matcher), A11 (Constructor), A13 (Renderer)
• LLM Required: No

Inputs

Input	Source	Format
PDF file	User input	File path
Paper markdown	workspace/markdown.md	Plain text

Execution Steps

Step 1: Extract Figures

1. Extract all figure images from the PDF
2. Save each as workspace/figures/image_{N}.png (numbered from 1)
3. For each image, record metadata:
   • caption: Figure caption text (from PDF or nearby text)
   • page_no: Page number where figure appears
   • image_path: Relative path to saved image file
   • width, height: Image dimensions in pixels
   • figure_aspect: width / height ratio

Step 2: Extract Tables

1. Identify all tables in the PDF
2. For each table, save a rendered image: workspace/figures/table_{N}.png
3. Record metadata:
   • caption: Table caption text
   • page_no: Page number
   • table_path: Relative path to saved table image
   • width, height: Dimensions

Step 3: Extract References

1. Find the "References" section in workspace/markdown.md
2. Parse into two dictionaries:
   • numbered: Maps reference numbers to full citation text

     {"1": "Smith et al., Deep Learning...", "2": "Jones, 2023..."}
     

   • citations: Maps "Author et al., Year" patterns to full text

     {"Smith et al., 2022": "Smith et al., Deep Learning..."}
     

Output Schema

workspace/images.json

{
  "image_1": {
    "caption": "Figure 1: System overview...",
    "page_no": 3,
    "image_path": "workspace/figures/image_1.png",
    "width": 800,
    "height": 600,
    "figure_aspect": 1.33
  }
}