Jev 1.14: Object Detection, Document Verification, and Image Analysis

TypeSafe has released Jev 1.14, introducing experimental support for image processing, a new 'box' query type, and expanded context capabilities. The author conducted a series of practical experiments, including animal breed classification, object detection based on descriptions, and extracting banking details from documents. A key focus was the challenge of identifying a specific object within a large set of photos. The results showed that model performance is highly sensitive to query structure: when processing 96 images, the model initially failed, but achieved perfect accuracy after the prompt structure was optimized. The total cost of these experiments was approximately $2.6. The article highlights the importance of prompt engineering for multimodal models, demonstrating how minor adjustments in data structure can significantly impact analysis results and prevent misinterpretation of a neural network's actual capabilities.