Illegible Region Detection Prompt for Scanned PDFs

This prompt is designed for document ingestion engineers and archival teams who need to programmatically assess the quality of scanned PDFs before text extraction. The core job-to-be-done is not just to read text, but to produce a machine-readable damage report: identifying specific regions of a page that are illegible, classifying the type of degradation (such as heavy blur, bleed-through from the reverse side, faint pencil marks, or physical obstructions like staples), and attaching a calibrated confidence score to each region. The ideal user is building an automated pipeline where silent OCR failures are unacceptable, and they need a triage step to route degraded pages for manual review or specialized image preprocessing.

You should use this prompt when you are processing large volumes of scanned archival material, legal discovery documents, or historical records where image quality is inconsistent. It is appropriate when your downstream system requires structured metadata about page quality to make decisions—for example, skipping text extraction on a page flagged as 90% illegible, or flagging a specific clause area for human transcription. The prompt requires a pre-processed input: a high-resolution page image and, optionally, the raw OCR text output for comparison. It is not a general-purpose OCR engine and should not be used as the primary text extraction step. It is a diagnostic layer that sits before extraction.

Do not use this prompt for clean, born-digital PDFs where text is already selectable and rendering artifacts are minimal. It is overkill for documents with uniform, high-contrast text. Avoid using it as a standalone quality gate without human review in high-stakes scenarios, such as single-source financial audits or medical record digitization, where a false negative (marking a region as legible when it is not) could have serious consequences. In those cases, the output should feed a human-in-the-loop review queue, not an automated deletion or acceptance workflow. The prompt also assumes a single page image per call; multi-page logic should be handled by your application's iteration harness.

This prompt template is designed to be dropped into a document processing pipeline that handles scanned PDFs. It expects pre-extracted text and layout metadata from an OCR engine as input, and it returns a structured JSON report of illegible regions. The template uses square-bracket placeholders for all variable inputs, constraints, and output schemas so you can adapt it to your specific OCR engine, document types, and downstream requirements without rewriting the core instruction logic.

textCopy
You are an expert document quality analyzer. Your task is to examine the provided OCR output and page metadata from a scanned PDF page and identify all regions where the text is illegible or likely to be incorrectly transcribed.

## INPUT
- OCR Text: [OCR_TEXT]
- OCR Confidence Scores (word-level): [OCR_CONFIDENCE_MAP]
- Page Layout Data (bounding boxes): [PAGE_LAYOUT]
- Page Image Description (optional): [PAGE_IMAGE_DESCRIPTION]

## TASK
1. Identify all regions on the page where the text is likely illegible, unreadable, or incorrectly transcribed.
2. For each identified region, classify the primary degradation type.
3. Estimate a confidence score (0.0 to 1.0) for the OCR output in that region.
4. Provide the bounding box coordinates for each illegible region.
5. If possible, suggest what the illegible text might be based on context, but clearly mark this as a guess.

## DEGRADATION TYPES
Classify each region into one of the following types:
- `blur`: Out-of-focus or motion-blurred text
- `bleed_through`: Text from the reverse side of the page showing through
- `faint_text`: Low-contrast or faded text
- `obstruction`: Text covered by a stain, stamp, mark, or physical object
- `skew_warp`: Text distorted by page skew or warping
- `font_degradation`: Broken, compressed, or stylized fonts that OCR cannot parse
- `background_noise`: Text obscured by background patterns or artifacts
- `handwriting`: Cursive or handwritten text that printed OCR cannot reliably transcribe
- `other`: Any other degradation not listed above (describe briefly)

## OUTPUT SCHEMA
Return a JSON object conforming to this exact schema:
{
  "page_number": [PAGE_NUMBER],
  "overall_page_quality": "high" | "medium" | "low" | "unusable",
  "illegible_regions": [
    {
      "region_id": "string",
      "bounding_box": {
        "x": number,
        "y": number,
        "width": number,
        "height": number,
        "unit": "pixel" | "point" | "inch"
      },
      "degradation_type": "blur" | "bleed_through" | "faint_text" | "obstruction" | "skew_warp" | "font_degradation" | "background_noise" | "handwriting" | "other",
      "degradation_description": "string (brief description of the degradation)",
      "ocr_confidence_estimate": number (0.0 to 1.0),
      "affected_text_guess": "string or null (your best guess at the text, or null if impossible)",
      "recommended_action": "re_scan" | "manual_review" | "alternate_ocr_engine" | "image_preprocessing" | "ignore"
    }
  ],
  "processing_notes": "string (any additional observations about the page quality)"
}

## CONSTRAINTS
- Only report regions that are genuinely illegible. Do not flag regions with minor noise that a human could easily read.
- If the entire page is clean and legible, return an empty `illegible_regions` array and set `overall_page_quality` to "high".
- If you cannot determine the degradation type with confidence, use "other" and describe what you observe.
- Do not invent text guesses. If you cannot make a reasonable guess based on context, set `affected_text_guess` to null.
- Bounding box coordinates must correspond to the coordinate system specified in [PAGE_LAYOUT].

## EXAMPLES
[EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]

To adapt this template, replace each square-bracket placeholder with your pipeline's actual data and configuration. The [OCR_TEXT] and [OCR_CONFIDENCE_MAP] placeholders should receive output directly from your OCR engine (such as Tesseract, Azure Form Recognizer, or Amazon Textract). The [PAGE_LAYOUT] placeholder expects structured bounding box data—pass it in the coordinate system your downstream system uses. The [EXAMPLES] placeholder is critical for few-shot performance; include at least two annotated examples showing both a degraded page and a clean page. Set [RISK_LEVEL] to guide the model's caution: use "low" for archival digitization where false positives are acceptable, "high" for legal or medical documents where missed illegible regions could cause harm, and "medium" for most enterprise workflows. For high-risk use cases, always route flagged regions to human review before the extracted text enters any downstream system.

The illegible region detection prompt is designed to operate as a post-OCR quality gate within a document ingestion pipeline. After a scanned PDF passes through an OCR engine, each page image and its corresponding OCR text output are sent to the model. The prompt returns a structured JSON payload identifying regions where the OCR confidence is low, classifying the degradation type (blur, bleed-through, faint text, obstruction), and providing bounding box coordinates for each illegible zone. This output should be consumed by a downstream routing decision: pages with no flagged regions proceed to extraction; pages with flagged regions above a configurable threshold are routed to a human review queue or a higher-resolution re-scan workflow.

Wire the prompt into your application as a stateless function that accepts a page image (base64-encoded) and the raw OCR text for that page. The function should enforce a strict output schema using a JSON validator before the result is accepted. If the model returns malformed JSON, missing required fields (regions, page_confidence), or coordinates outside the page dimensions, the application should retry once with a repair prompt that includes the validation error message. After two consecutive failures, escalate the page to human review and log the failure for prompt debugging. For model choice, use a vision-capable model (GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro) since the task requires comparing the rendered page image against the OCR text. Set temperature=0 to minimize coordinate drift and classification inconsistency across runs.

Build an evaluation harness before deploying to production. Create a golden dataset of 50–100 scanned pages with known degradation types and manually annotated illegible region bounding boxes. For each page, run the prompt and compare the model's output against ground truth using Intersection over Union (IoU) for region localization and exact-match accuracy for degradation type classification. Set acceptance thresholds: IoU ≥ 0.7 for region detection and ≥ 85% accuracy for degradation classification. Run this eval suite on every prompt change and on a weekly cadence against production samples to detect model behavior drift. Log every model response, including confidence scores and region counts, to monitor for sudden spikes in flagged regions that may indicate upstream scanner issues rather than prompt failures.

The highest-risk failure mode in production is a false negative—an illegible region that the model misses entirely. A missed illegible region propagates garbled text into downstream extraction, where it may produce silently corrupted structured data. Mitigate this by implementing a second-pass check: after extraction runs on pages that passed the illegibility gate, run a low-confidence extraction flagging prompt on the extracted fields. If the extraction prompt flags fields with low confidence that correspond to regions the illegibility prompt marked as clean, log the discrepancy and route the page for human review. This dual-gate pattern catches both detection failures and extraction failures before data enters your system of record. Do not rely on the illegibility prompt alone for high-stakes document processing; it is a filter, not a guarantee.