Multi-Modal Prior Art Searches
What is a Multi-Modal Search?
A multi-modal search relies on multiple forms of data, such as text and images, which is especially relevant for patents and scientific literature. By combining visual intelligence with traditional LLM-based models, multi-modal systems help AI systems "see" documents and extract a richer semantic understanding across text and figures.
Crafting Searches
At Palisade, we work with our clients to develop the most relevant and effective prior art search strategy. We analyze the prosecution history to determine the claim elements or arguments that overcame rejections. We are always available to discuss key elements with our clients or their outside counsel.
Collecting Results
A technical analyst will search for patents and NPL that may be relevant using a proprietary search engine, developed by Palisade, collecting materials from 15 patent offices and over 100 industry publications and scientific journals. Then, Palisade uses multi-modal AI models to prioritize the review of the most relevant prior art results.
Multi-Modal Analysis
Palisade uses a multi-modal embedding approach that revolutionizes document retrieval and understanding. Rather than relying on OCR or mechanical text extraction, Palisade feeds images of documents into a vision-language encoder trained via contrastive learning (similar to CLIP). Each page is broken into a grid and every patch is embedded into a shared semantic space.
The multi-modal embedding technique used by Palisade is especially powerful for patents, which often contain figures, tables, chemical structures, or other diagrams that don’t translate well into text. Traditional LLM-based retrieval systems depend on accurate text extraction and sentence-level transformation. In contrast, Palisade captures richer cross-modal semantic, making it more effective at identifying relevant figures and diagrams, leading to stronger prior art.