Adding Reference Examples

Adding Reference Examples

This is the most impactful contribution you can make to PaperBanana. Output quality improves directly with better reference examples.

What makes a good reference example

• The diagram clearly illustrates system architecture, pipeline flow, or framework structure
• Landscape layout with aspect ratio between 1.5 and 2.5 (width / height)
• The methodology section is self-contained enough to describe the approach without the full paper
• The diagram is not a results plot, ablation table, t-SNE visualization, or data sample

Option 1: Submit a paper recommendation

Lowest effort. Open a Reference Example issue with:

• arXiv link
• Figure number
• Which category it fits

We handle extraction and curation.

Option 2: Submit a complete reference via PR

Step 1: Extract the data

You need three things from the paper:

1. Methodology text: Copy the methodology/method section from the paper. Clean up any PDF extraction artifacts (broken words, stray headers). The text should describe what the diagram shows.

2. Diagram image: Extract the methodology figure as a PNG. Ensure it's clean, readable at 800px width, and has an aspect ratio between 1.5 and 2.5.

3. Metadata: Paper title, arXiv ID, figure number, original caption, and category.

Step 2: Create the directory

data/reference_sets/your_example_name/
├── methodology.txt
├── diagram.png
└── metadata.json

Use a short, descriptive directory name based on the paper's key concept (e.g., react_agent, diffusion_unet, graph_rag).

Step 3: Populate metadata.json

{
  "paper_title": "ReAct: Synergizing Reasoning and Acting in Language Models",
  "arxiv_id": "2210.03629",
  "figure_number": 1,
  "caption": "Overview of the ReAct framework combining reasoning traces and actions",
  "category": "agent_reasoning",
  "source_url": "https://arxiv.org/abs/2210.03629",
  "aspect_ratio": 2.1
}

Step 4: Verify

Before submitting the PR:

• The methodology text matches what the diagram actually depicts
• The diagram is clean (no scan artifacts, no cropping issues)
• The aspect ratio is between 1.5 and 2.5
• The paper is publicly available
• metadata.json is valid JSON with all fields populated

Step 5: Submit PR

Open a pull request with the title [Reference]: <paper name>. The PR template will guide you through the checklist.

Categories we need most

• Science & Applications: Domain-specific architectures outside core ML
• Vision & Perception: Detection, segmentation, multimodal pipelines

Any category is welcome, but these two are currently underrepresented.