AI Safety Filters

Why Safety Matters

PDFs can contain hidden content that’s invisible to humans but extracted by parsers. Malicious PDFs can use this to inject prompts into LLM pipelines.

Safety Filters

EdgeParse includes built-in content safety filters:

Hidden Text Detection

Text rendered with the same color as the background, or with render mode 3 (invisible), is flagged and optionally excluded.

Off-Page Content

Content positioned outside the visible page area (negative coordinates, beyond page boundaries) is detected and filtered.

Tiny Text

Text smaller than a configurable threshold (default: 3pt) is flagged. This catches hidden micro-text used for SEO manipulation or prompt injection.

OCG Layer Filtering

Optional Content Groups (OCG) that are set to invisible by default are detected and handled appropriately.

Configuration

import edgeparse

# Safety filters are enabled by default
markdown = edgeparse.convert("document.pdf", format="markdown")

# All content including hidden text
markdown = edgeparse.convert("document.pdf",
    format="markdown",
    include_hidden=True
)

RAG Security

When building RAG pipelines, these filters help prevent:

• Prompt injection — hidden instructions embedded in PDFs
• Data poisoning — invisible content that skews embeddings
• Context pollution — irrelevant hidden text diluting retrieval quality