Winnie-the-Pooh in 768 Dimensions: Semantic Search for Codex Pets on YDB

The Codex Pets catalog contains 153 approved animated pets that accompany AI agents. As the catalog grew, tag-based search became impractical, failing to find pets described by appearance rather than exact tags. The author implemented semantic search: text fields (name, description, tags) are encoded into 768-dimensional vectors, and a multimodal model generates captions for four selected animation frames which are also vectorized. YDB stores the vectors and computes cosine similarity via Knn::CosineSimilarity. Four new tables were added to handle embeddings, captions, recommendations, and versioning. The search merges three ranked lists using Weighted Reciprocal Rank Fusion, with lexical and textual weights of 1 and a tuned visual weight. The visual branch can fail independently, falling back to lexical results. Testing showed semantic search improves NDCG@5 from 0.539 to 0.724 and Recall@5 from 0.528 to 0.743 on the full set of 14 queries.