Marketing banners require structured OCR retrieval

Marketing images are visual-first, making traditional OCR-to-text pipelines brittle; success requires a modular architecture that prioritizes layout over raw strings.

• –Layout-aware detection via PP-StructureV3 is critical for separating conflicting context like headlines and fine print.
• –PP-ChatOCRv4 provides a robust local path for zero-shot field extraction (Price, Promo Code) without cloud dependencies.
• –Markdown output is the superior intermediate format for RAG, as it preserves the hierarchy LLMs need to interpret semi-structured data.
• –Validation layers using Pydantic or similar schemas are essential for sanitizing OCR noise before it hits the retrieval layer.
• –Hybrid retrieval combining CLIP visual features with semantic text embeddings covers gaps where stylized fonts defeat standard character recognition.