Blogs›Weaviate›Semantic Search for Creative Assets
This post introduces the concept of applying semantic search and vector databases to creative asset management, addressing the challenges of retrieval and organization in creative workflows. It explains how AI can be used as a workflow layer to make existing work more accessible by converting media into vector embeddings for retrieval based on meaning rather than keywords. The post outlines the ingestion process and the role of vector databases in storing embeddings alongside original content. This post extends this to multimodal embeddings, enabling search across text, images, audio, and video by mapping them into a shared embedding space. It details the use of natively multimodal embedding models like Gemini Embedding 2 and discusses design decisions such as native vs. bridge-based embedding, chunking strategies for non-text data, and dimension size with Matryoshka Representation Learning (MRL). It provides practical examples of building multimodal systems using Weaviate, Gemini Embedding 2, and Gemini 3 Flash for searching audio without a transcript, searching images with text queries, and performing multimodal RAG with video.
7/30/2026
Introduces the application of semantic search and vector databases to creative asset management. Explains the process of converting media into vector embeddings using ML models and how vector search enables retrieval by meaning. Highlights the importance of metadata alongside embeddings and the role of vector databases. Discusses the ingestion layer challenges and previews a future implementation.
6/18/2026
Introduces server-side batching for efficient data import, a retry-with-checkpoint pattern for robust error handling, and the `blobHash` data type for storing media embeddings while omitting raw bytes. Also discusses multimodal ingestion for PDFs and images using Weaviate Embeddings and `multi2vec-google`.
5/6/2026
This post details the critical role of retrieval quality in Retrieval-Augmented Generation (RAG) systems and its impact on LLM output. It identifies five specific retrieval failure modes: retrieval drift, context truncation, stale index poisoning, low-relevance top-k retrieval, and inter-agent miscommunication. It proposes four key areas for improving retrieval quality: embedding model selection, chunking architecture, retrieval strategy (including hybrid search, cross-encoder re-ranking, and relevance thresholding), and index maintenance and freshness. Finally, it outlines a practical measurement framework for evaluating retrieval quality using metrics like context precision and context recall.
4/1/2026
Introduces and details the practical application of multimodal embeddings for semantic search and RAG. Explains how natively multimodal embedding models (like Gemini Embedding 2) map text, images, audio, and video into a shared embedding space. Discusses key design decisions: native vs. bridge-based embedding, chunking strategies for non-text data (fixed-width time windows for audio/video, page-level for documents), and the use of Matryoshka Representation Learning (MRL) for dimension reduction. Provides concrete examples of building multimodal systems with Weaviate, Gemini Embedding 2, and Gemini 3 Flash, including searching audio without transcripts, searching images with text, and multimodal RAG with video.