Temenuzhka Dimova
The 2025 revisited edition of the Vasari Project moves beyond static graph structures to embrace a new paradigm of semantic exploration in art history. By integrating Weaviate, a vector-based search engine, the project now enables contextual and conceptual connections across vast corpora of Wikipedia articles sections ( with a total of 445,668 data objects indexed ) .
Instead of relying solely on hyperlink structures, this new version leverages AI-driven embeddings to detect deep relationships between historical figures, artistic movements, iconographic themes, and institutions. Through semantic search, users can now query not just "who influenced whom", but "how artistic intentions, political gestures, and gendered narratives echo across centuries."
The interactive interface dynamically surfaces multi-layered relations: painters and non-painters, curators, places, techniques, artworks, and thematic categories — all projected in a visually intuitive circular diagram. These graphs are enriched by machine-learned associations, allowing for unexpected cross-temporal insights, such as the conceptual link between a 14th-century altarpiece and a 20th-century performance art piece.
At its core, Weaviate enables the indexing and retrieval of high-dimensional vectors. Each object—such as a Wikipedia section, an image caption, or a textual description—is transformed into a numeric vector and stored within a vector index. Upon receiving a search query, Weaviate computes the query's vector and performs nearest-neighbor search to retrieve semantically similar items based on distance metrics like cosine similarity.
One of Weaviate’s distinguishing features is its ability to support named vectors and multi-vector configurations, enabling different fields (e.g., title, abstract, content) to be indexed using specialized models. This is particularly advantageous in domains where vocabulary is heterogeneous or where concepts are abstract and context-dependent, such as art history or the humanities.
To further enhance the quality of results, Weaviate supports re-ranking mechanisms. These apply a second-stage model—such as Cohere’s rerank-english-v3.0—to reorder initial results based on contextual relevance. Such methods are crucial when the initial vector search returns semantically plausible, but not contextually precise, matches.
Weaviate also accommodates hybrid retrieval, combining classical keyword filtering (BM25-style) with semantic similarity, thereby ensuring both precision and recall in complex search tasks.