Traditionally, implementing these capabilities required specialized computer vision infrastructure, external vector databases, custom ML pipelines, and multiple frameworks.<\/p>\n\n\n\n
With MySQL HeatWave GenAI<\/a><\/strong>, many of these capabilities can now be implemented directly inside SQL workflows<\/strong> using built-in AI routines such as:<\/p>\n\n\n\n In this article, we will build the foundations of a Vision Model Evaluation Assistant<\/strong> using MySQL HeatWave GenAI<\/strong>.<\/p>\n\n\n\n The objective is not to create another image classifier, but rather a semantic image understanding platform<\/strong> capable of:<\/p>\n\n\n\n This article intentionally focuses on the database layer and the SQL logic powering the system.<\/p>\n\n\n\n Before building the solution, it is important to understand the major approaches currently used for AI-powered image understanding systems.<\/p>\n\n\n\n Let’s have a look on three paradigms for representing images in AI systems.<\/p>\n\n\n\n The first approach consists of converting raw image pixels directly into vectors.<\/p>\n\n\n\n Popular architectures include:<\/p>\n\n\n\n In this approach:<\/p>\n\n\n\n These embeddings are often difficult to interpret:<\/p>\n\n\n\n If a search result is incorrect, it is difficult to understand why<\/em> the model failed.<\/p>\n\n\n\n This is the approach used in this article.<\/p>\n\n\n\n Instead of embedding raw pixels directly, we first generate a textual understanding of the image.<\/p>\n\n\n\n The workflow becomes:<\/p>\n\n\n\n This pipeline is composed of two distinct stages:<\/p>\n\n\n\n A Vision Language Model<\/a> (VLM) analyzes the image and generates:<\/p>\n\n\n\n Example<\/span>:<\/p>\n\n\n\n The generated text is then converted into a vector embedding using a text embedding model.<\/p>\n\n\n\n This creates semantic vectors representing the meaning of the image.<\/p>\n\n\n\n This approach offers several advantages:<\/p>\n\n\n\n Unlike direct image embeddings, the generated metadata is understandable by humans.<\/p>\n\n\n\n If a search fails, developers can inspect:<\/p>\n\n\n\n This creates a powerful audit trail.<\/p>\n\n\n\n Because the representation becomes textual:<\/p>\n\n\n\n This aligns perfectly with MySQL HeatWave GenAI capabilities<\/strong>.<\/p>\n\n\n\n The quality of retrieval depends heavily on:<\/p>\n\n\n\n Low-level visual details may also be lost:<\/p>\n\n\n\n A third category consists of multimodal architectures capable of jointly processing:<\/p>\n\n\n\n These systems use:<\/p>\n\n\n\n Their objective is to create unified multimodal embeddings.<\/p>\n\n\n\n These architectures are often:<\/p>\n\n\n\n In this article, we intentionally choose the Semantic Metadata Embeddings<\/strong> pipeline because it aligns perfectly with MySQL HeatWave GenAI capabilities.<\/p>\n\n\n\n Specifically<\/span><\/strong>:<\/p>\n\n\n\n This approach also creates a human-readable semantic audit trail.<\/p>\n\n\n\n If retrieval quality degrades, architects and developers can directly inspect the generated descriptions to understand:<\/p>\n\n\n\n For AI evaluation systems, explainability is often just as important as accuracy<\/strong>.<\/p>\n\n\n\n\n
sys.ML_GENERATE<\/a><\/code><\/li>\n\n\n\nsys.ML_EMBED_ROW<\/a><\/code><\/li>\n\n\n\nDISTANCE<\/a><\/code> and the other Vector functions<\/a>.<\/li>\n<\/ul>\n\n\n\n\n
Landscape of Vision-to-Embedding Approaches<\/h1>\n\n\n\n
1. Direct Image Embeddings<\/h2>\n\n\n\n
\n
\n
Workflow<\/h3>\n\n\n\n
![\"\"](\"https:\/\/i0.wp.com\/dasini.net\/blog\/wp-content\/uploads\/2026\/05\/1.1.workflow_Direct_Image_Embedding.png?resize=620%2C160&ssl=1\")
<\/figure>\n\n\n\nStrengths<\/h3>\n\n\n\n
\n
Limitations<\/h3>\n\n\n\n
\n
To summarize<\/h3>\n\n\n\n
Pros<\/h4>\n\n\n\n
\n
Cons<\/h4>\n\n\n\n
\n
2. Semantic Metadata Embeddings<\/h2>\n\n\n\n
![\"\"](\"https:\/\/i0.wp.com\/dasini.net\/blog\/wp-content\/uploads\/2026\/05\/1.2workflow_Semantic_Metadata_Embeddings-1.png?resize=800%2C206&ssl=1\")
<\/figure>\n\n\n\nStep 1: Image -> <\/em>Text<\/h3>\n\n\n\n
\n
{<\/code> \"description\": \"A red motorcycle parked on a city street near buildings during daytime.\",\n\n \"keywords\": [\"motorcycle\", \"street\", \"urban\", \"vehicle\", \"city\"]\n}<\/code><\/code><\/pre>\n\n\n\nStep 2: Text -> <\/em>Embedding<\/h3>\n\n\n\n
Strengths<\/h3>\n\n\n\n
Human-readable and explainable<\/h4>\n\n\n\n
\n
SQL-native workflow<\/h4>\n\n\n\n
\n
Trade-offs<\/h3>\n\n\n\n
\n
\n
To summarize<\/h3>\n\n\n\n
Pros<\/h4>\n\n\n\n
\n
Cons<\/h4>\n\n\n\n
\n
3. Multimodal Fusion Models<\/h2>\n\n\n\n
\n
\n
Strengths<\/h3>\n\n\n\n
\n
Challenges<\/h3>\n\n\n\n
\n
Why We Chose the Image -> Text -> Embedding Approach?<\/h2>\n\n\n\n
\n
ML_GENERATE<\/strong><\/a><\/code><\/code> handles image-to-text generation<\/li>\n\n\n\nML_EMBED_ROW<\/strong><\/a><\/code><\/code> handles text-to-vector conversion<\/li>\n<\/ul>\n\n\n\n\n
Overview of MySQL HeatWave GenAI<\/h1>\n\n\n
![\"MySQL](\"https:\/\/i0.wp.com\/dasini.net\/blog\/wp-content\/uploads\/2025\/09\/rc24-heatwave-logo.gif?fit=770%2C550&ssl=1\"){kind=logo}