Frank Liu is the Voyage AI research manager at MongoDB, working on the infrastructure that helps AI agents retrieve accurate information. His career spans computer vision, startup hardware, vector databases, and the embedding models behind modern semantic search.
Liu earned bachelor’s and master’s degrees in electrical engineering at Stanford, where his graduate research explored how machine learning could assess image aesthetics. He began his career working on computer vision and machine learning at Yahoo, then co-founded Orion Innovations, a Shanghai indoor-positioning startup where, as chief technology officer, he worked across custom hardware, machine learning, cloud software, and visualization.
At Zilliz, Liu held engineering and operations leadership positions, contributed to the Milvus ecosystem, and maintained Towhee, an open-source embedding framework for turning unstructured data into searchable representations. His writing demystifies the mechanics underlying that work, from vector databases and semantic search to hierarchical navigable small-world indexes.
He subsequently joined Voyage AI and moved to MongoDB following its acquisition of the embedding-model company. His responsibilities have included product management; more recently, he has led Voyage AI research at MongoDB.
- Retrieval requires more than vectors. Liu emphasizes embeddings, reranking, structured filters, and workload-specific evaluation. Code search benefits from models attuned to programming and documentation; legal retrieval also needs constraints such as jurisdiction and document type. His AI Engineer conference talk frames these choices as application requirements, not interchangeable infrastructure.
- Agentic retrieval is iterative. Agents can decompose complicated questions into multiple searches, expand queries, incorporate conversational context, and refine their answers as new information arrives. Retrieval accuracy matters because errors can compound through subsequent decisions.
- Multimodal search should reflect real enterprise data. Liu advocates embedding text, images, and mixed-media documents into a shared semantic space, alongside instruction-aware retrieval that directs results toward specific user needs.
- Databases should eliminate retrieval plumbing. His work on production-ready agent retrieval describes automated Voyage AI embeddings in Atlas, embedding and reranking APIs, the voyage-code-4 model, streaming vector search, and coordination between MongoDB and Temporal for agent state and changing operational data.
