Callout Box and Sidebar Content Isolation Prompt Template

This prompt is designed for document ingestion engineers who need to separate visually distinct content regions—such as callout boxes, sidebars, pull quotes, and info boxes—from the main body text of a document. The core job-to-be-done is preventing layout misordering from silently corrupting downstream retrieval, chunking, and structured extraction pipelines. The ideal user has already run a layout detection model (producing bounding boxes, font metadata, and text spans) and needs a reliable, typed extraction of these regions before the text is linearized for RAG or database ingestion. You should use this prompt when your document set includes textbooks, technical manuals, research papers, legal filings, or marketing collateral where sidebars contain definitions, warnings, or supplementary data that must not be interleaved with the narrative flow.

Do not use this prompt as a substitute for a layout detection model. It expects structured layout metadata as input—bounding box coordinates, font sizes, style tags, and page dimensions—not raw PDF bytes or plain text. It is also not the right tool for extracting the main reading order; pair it with a multi-column reading order recovery prompt to linearize the body text after isolation. The prompt is optimized for documents where callout regions are visually distinct through background shading, border boxes, font-size deltas, or margin offsets. If your documents use only inline bold or italic text for emphasis without spatial separation, a simpler inline-style detection prompt will be more appropriate. For high-stakes legal or compliance workflows, always route the extracted regions through human review before accepting them as ground truth for downstream actions.

Before wiring this into production, prepare a small golden dataset of 10–20 pages with known callout and sidebar regions. Run the prompt and measure two failure modes: body text contamination inside isolated regions, and callout text leaking into the body stream. If either rate exceeds 5%, adjust the spatial proximity thresholds in the prompt's constraints or add a post-extraction validator that checks for semantic continuity breaks at region boundaries. For documents with scanned pages, ensure OCR confidence scores are passed alongside text spans so the prompt can flag low-certainty regions for manual review instead of silently misclassifying them.

This prompt template isolates callout boxes, sidebars, pull quotes, info boxes, and other visually distinct content regions from body text. It expects pre-extracted layout data—bounding boxes, font metadata, and text spans—rather than raw PDF bytes. The model's job is to classify each region as body text or a distinct callout type, then produce typed region objects with content and position metadata. Use this when your ingestion pipeline already has layout extraction but needs semantic region classification before chunking or structured extraction.

codeCopy
You are a document layout classifier. Your task is to identify and isolate callout boxes, sidebars, pull quotes, info boxes, and other visually distinct content regions from body text in a document page.

## INPUT
You will receive layout data for a single document page as a JSON array of text regions. Each region includes:
- `region_id`: unique identifier
- `bbox`: [x0, y0, x1, y1] in page coordinates
- `text`: the text content of the region
- `font_name`: primary font used
- `font_size`: primary font size in points
- `font_color`: hex color string or null
- `has_background`: boolean indicating background shading or fill
- `has_border`: boolean indicating visible border or stroke

[LAYOUT_DATA]

## TASK
For each region, classify it as one of the following types:
- `body_text`: standard flowing text in the main content column
- `callout_box`: a boxed or shaded region containing emphasized content, often with a border or background
- `sidebar`: a column or block parallel to the main text flow, typically narrower and positioned at page edges
- `pull_quote`: a large, stylized quotation extracted from body text, often with distinct font size and style
- `info_box`: a bordered or shaded region containing supplementary facts, definitions, or notes
- `header`: running header text repeated across pages
- `footer`: running footer text including page numbers
- `caption`: figure, table, or image caption text
- `footnote`: bottom-of-page note with or without a reference marker
- `other`: any region that does not fit the above categories

## CLASSIFICATION RULES
1. Use position: sidebars appear at page edges outside the main text column. Callout boxes interrupt or sit beside body text.
2. Use visual cues: borders, background shading, and distinct font sizes signal callout boxes and info boxes.
3. Pull quotes use significantly larger or stylized fonts and repeat content found elsewhere.
4. Headers and footers repeat across pages with consistent position near page top or bottom.
5. When uncertain between `callout_box` and `body_text`, prefer `callout_box` if the region has a border or background.
6. Do not split a single region into multiple types.

## OUTPUT SCHEMA
Return a JSON object with this exact structure:
```json
{
  "page_regions": [
    {
      "region_id": "string",
      "type": "callout_box|sidebar|pull_quote|info_box|body_text|header|footer|caption|footnote|other",
      "content": "full text of the region",
      "bbox": [x0, y0, x1, y1],
      "confidence": 0.0_to_1.0,
      "evidence": ["list of cues used for classification, e.g., 'border present', 'position at right edge', 'font_size 18pt vs body 11pt'"]
    }
  ],
  "body_text_contamination_warnings": [
    {
      "callout_region_id": "string",
      "warning": "description of body text that may have been incorrectly merged into this callout region",
      "severity": "low|medium|high"
    }
  ]
}

CONSTRAINTS

• Classify every input region. Do not skip regions.
• Set confidence below 0.7 when visual cues are ambiguous or conflicting.
• Flag potential body text contamination in callout regions when text length or style seems inconsistent with a standalone callout.
• If [RISK_LEVEL] is "high", include a requires_human_review boolean field set to true for any region with confidence below 0.8.

Adapt this template by replacing [LAYOUT_DATA] with your extraction pipeline's output. If your layout data includes additional fields such as column_index or reading_order, add them to the input schema and reference them in the classification rules. For high-stakes workflows such as legal or compliance documents, set [RISK_LEVEL] to "high" to enable the human-review flag. Wire the output into a downstream validator that checks for body text contamination warnings before the isolated regions enter your chunking or structured extraction pipeline. The next section covers how to integrate this prompt into an application harness with validation, retries, and logging.

This prompt is designed to operate as a post-layout-extraction processing step, not as a standalone PDF reader. The input to this prompt should already contain bounding box coordinates, font metadata, and text spans from a PDF parser such as PyMuPDF, pdfplumber, or a cloud document AI service. The prompt expects a structured representation of page regions—typically a JSON array of text blocks with position, style, and content fields—rather than raw PDF bytes or plain text. This separation of concerns keeps the prompt focused on classification and isolation logic while the upstream parser handles rendering and coordinate extraction.

Wire this prompt into your pipeline after layout extraction but before chunking or structured extraction. The typical flow: (1) PDF parser produces text blocks with bounding boxes and font metadata, (2) this prompt classifies each block as body, callout, sidebar, pull quote, or info box, (3) a post-processing validator checks that body text regions contain no orphaned callout content, (4) isolated regions are routed to separate extraction paths or tagged with region-type metadata for downstream consumers. Implement a retry loop with exponential backoff for malformed JSON responses—if the model returns invalid JSON or missing required fields, retry up to 3 times with the error message appended to the prompt. Log all classification decisions with block IDs and confidence scores for debugging layout misattribution issues later.

For production deployment, add a validation layer that checks for body text contamination—the most common failure mode where callout content bleeds into adjacent body paragraphs. Implement a cross-reference check: every text block classified as callout or sidebar must have its content verified against the body text stream to ensure no substring overlap exceeding 80% similarity. If contamination is detected, flag the page for human review or route to a secondary model with stricter isolation instructions. Use a model with strong JSON output discipline (GPT-4o, Claude 3.5 Sonnet, or equivalent) and set temperature=0 to minimize classification drift. For high-throughput pipelines, batch pages but keep each page's regions in a single prompt call to preserve spatial context. Store the raw prompt, model response, and validation results in your observability system for regression testing when the prompt or model version changes.