Different types of index structures have been devised to back different types of search queries. For example:
- Lexical search: Typically backed by an inverted index. We express queries using keywords or Boolean expressions, e.g., "documents that match 'brent crude' and that are published after 1985 and authored by people whose names contain the string 'john'."
- Semantic search: Typically backed by an approximate nearest-neighbour index. We express queries in natural language and use similarities between embedding vectors to assess fit, e.g., "documents that have to do with oil and gas exploration in the North Sea."
- Structural search: Typically backed by a graph index. We express queries with constraints on nodes and edges in a knowledge graph, e.g., "documents (nodes) representing people whose first name is 'Peter' and that are connected to other people whose last name is 'Pan' if we follow the relation (edges) named 'is-friends-with' for at most 2 hops."
In many applications, we want to express queries that span more than one paradigm, e.g., "documents that have to do with oil and gas exploration in the North Sea and that are authored by people whose names contain the string 'john'." One approach is to evaluate different parts of the query across independent systems and indices and post-process their results, e.g., via result fusion. More recently, native hybrid search systems have emerged to combine multiple paradigms within a unified framework, balancing expressive precision, latency, and operational efficiency.
In this master's thesis, the student will:
- Assess the state of the art in systems and query execution strategies that combine multiple search paradigms.
- Analyze the trade-offs between different architectures (e.g., late-fusion pipelines versus integrated hybrid indices) regarding latency, scalability, and result quality.
- Design and run benchmarking experiments to evaluate existing approaches and propose extensions or improvements to current methods.