> ## Documentation Index
> Fetch the complete documentation index at: https://docs.knowledgestack.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# PDF Watermark Removal

> The ingestion pipeline automatically removes watermarks from PDFs before conversion. Watermarks degrade OCR quality and pollute extracted text, so stripping them early improves downstream chunking ...

## When It Runs

Watermark removal runs during the document preparation step. If watermarks are detected, the cleaned PDF is uploaded to storage alongside the original; otherwise the original passes through unchanged.

Two complementary strategies are applied sequentially -- if either detects watermarks, the cleaned document is used for subsequent processing.

## Strategy 1: Annotation Watermarks

Some PDF producers (especially office-suite "confidential" or "draft" stamp features) add watermarks as annotation overlays on top of page content. The PDF specification defines a dedicated watermark annotation subtype for this purpose.

**How it works:** The system scans each page for annotations marked as watermarks and removes them. Because these are explicitly declared as watermarks by the producing application, no heuristics are needed -- the annotation type is a definitive signal.

**Limitation:** Many real-world watermarks are not implemented as annotations. Text-based watermarks embedded directly in the page content (common in documents from Chinese recruitment platforms, for example) are invisible to this strategy, which is why Strategy 2 exists.

## Strategy 2: Content-Stream Watermarks

This strategy targets watermarks embedded directly in the PDF page content as rotated, semi-transparent text. These are common in documents where the watermark was "baked in" during PDF generation rather than added as an overlay.

### Detection Heuristics

A text group on the page is classified as a watermark when **both** conditions are met:

| Condition            | Threshold                                |
| -------------------- | ---------------------------------------- |
| **Rotated text**     | Rotation angle between 25 and 75 degrees |
| **Semi-transparent** | Fill opacity below approximately 29%     |

Both conditions must be satisfied simultaneously. This means:

* Rotated but fully opaque text (e.g., decorative headings) is left untouched
* Horizontal but transparent text (e.g., background labels) is left untouched
* Only text that is both significantly rotated AND significantly transparent is removed

### How It Works

1. Each page's content stream is normalized into a single, well-formed structure
2. Text groups (graphics-state blocks) are analyzed for rotation angle and transparency level
3. Groups matching both criteria are removed from the content stream
4. The modified page content is written back

### Why This Approach?

The system operates on raw PDF drawing operations (rotation matrices and transparency settings) rather than trying to match specific watermark text. This makes it:

* **Encoding-independent** -- works regardless of font, language, or text encoding
* **Low false-positive risk** -- the combination of non-trivial rotation AND significant transparency is highly specific to watermarks; normal document text is almost always horizontal and fully opaque
* **Non-destructive** -- only the identified watermark elements are removed; all other page content is preserved

## Safety

* **False-positive risk is low**: Normal document text is horizontal (0 degrees) and opaque. The dual-condition requirement (rotated + transparent) is highly specific to watermarks.
* **Clean removal**: Watermark text groups are self-contained in the PDF structure, so removing them does not affect surrounding content.
* **Original preserved**: The original uploaded PDF is always kept in storage. The cleaned version is stored separately.