gitea-admin/gh-christianlouis-docuelevate
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(ocr): fine-tune OCR quality criteria with stricter threshold and head-to-head comparison

Browse Source 
- Raise quality acceptance threshold from 65→85 (configurable via TEXT_QUALITY_THRESHOLD)
- Reject text with significant issues (excessive_typos, garbage_characters,
  incoherent_text, fragmented_sentences) even when score is above threshold
  (configurable via TEXT_QUALITY_SIGNIFICANT_ISSUES)
- Add compare_text_quality() for AI-powered head-to-head comparison of
  original embedded text vs fresh OCR output
- Update process_document to pass original text to OCR task for comparison
- Update process_with_ocr to run comparison and keep the higher-quality text
- Add new settings to settings_service.py metadata
- Update docs/ConfigurationGuide.md with new settings
- Add comprehensive tests for new threshold and comparison logic

Co-authored-by: christianlouis <361235+christianlouis@users.noreply.github.com>
This commit is contained in:
copilot-swe-agent[bot]
2026-02-25 13:03:21 +00:00
parent d1c60c5ac5
commit 89d5df71c8
7 changed files with 659 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/ConfigurationGuide.md
+12 -5
View File
@@ -465,9 +465,11 @@ When multiple providers are listed, all run in parallel and their results are me
DocuElevate can automatically assess whether the text already embedded in a PDF is of sufficient quality before deciding to skip OCR. This prevents poor OCR output from a previous scan being silently used for downstream processing.
| **Variable** | **Description** | **Default** |
|--------------------------------|---------------------------------------------------------------------------------|-------------|
| `ENABLE_TEXT_QUALITY_CHECK` | Enable AI-based quality assessment of embedded PDF text. | `true` |
| **Variable** | **Description** | **Default** |
|-------------------------------------------|---------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------|
| `ENABLE_TEXT_QUALITY_CHECK` | Enable AI-based quality assessment of embedded PDF text. | `true` |
| `TEXT_QUALITY_THRESHOLD` | Minimum quality score (0–100) required to accept embedded text without re-OCR. | `85` |
| `TEXT_QUALITY_SIGNIFICANT_ISSUES` | Comma-separated issue labels that force re-OCR even when the score meets the threshold. | `excessive_typos,garbage_characters,incoherent_text,fragmented_sentences` |
**How it works:**
@@ -478,11 +480,16 @@ DocuElevate can automatically assess whether the text already embedded in a PDF
- Garbage characters or symbol soup
- Incoherent or nonsensical sentences
- Heavy fragmentation
4. If the quality score falls **below 65/100**, the embedded text is discarded and the file is sent to the configured OCR providers for a fresh scan.
5. All quality decisions (score, source, AI feedback) are recorded in the processing log for review.
4. The text is **rejected** (and re-OCR triggered) when **either** of these conditions is true:
- The quality score is **below** `TEXT_QUALITY_THRESHOLD` (default 85), **or**
- The AI identifies one or more issues listed in `TEXT_QUALITY_SIGNIFICANT_ISSUES` — even if the numeric score is above the threshold. This prevents edge cases such as a score of 68 with `excessive_typos` and `garbage_characters` being silently accepted.
5. After the re-OCR pass, the fresh OCR result is compared **head-to-head** against the original embedded text using an AI side-by-side review. The higher-quality text is passed to downstream processing (metadata extraction, AI analysis). This ensures re-OCR never degrades quality.
6. All quality decisions (score, source, AI feedback, comparison outcome) are recorded in the processing log for review.
> **Tip**: Set `ENABLE_TEXT_QUALITY_CHECK=false` to disable the check entirely and always use embedded text as-is. This is useful when the AI provider is unavailable or when processing speed is more important than text accuracy.
> **Tuning the threshold**: The default of `TEXT_QUALITY_THRESHOLD=85` is intentionally strict. Lower it (e.g., `70`) for environments with consistently good existing OCR. Raise it (up to `100`) for maximum quality enforcement.
#### Searchable PDF Text Layer
Not all OCR providers embed a searchable text layer in the output PDF. The table below summarises each provider's behaviour and how DocuElevate handles it: