My paper "Semantic Caching for OLAP via LLM-Based Query Canonicalization" has been accepted at DOLAP 2026 (28th International Workshop on Design, Optimization, Languages and Analytical Processing of Big Data), co-located with EDBT/ICDT 2026 in Tampere, Finland.
This started as a project I had offered to interns and students, but nobody took me up on it. After talking with some collaborators who liked the idea, I decided to just do it myself. I will be presenting the paper at DOLAP in Tampere myself, so come check it out if you are interested.
The idea
As LLMs become the interface to databases, cache keys stop matching. Users ask the same analytical question in different ways, text-to-SQL systems produce different SQL for the same intent, and traditional caches that key on SQL text or AST miss all of it. Changing the interface breaks caching, and OLAP queries are too expensive to keep recomputing.
This paper brings caching back for dashboard-style OLAP workloads. It canonicalizes both SQL and natural language into a shared structured key, an OLAP Intent Signature, that captures what the query computes rather than how it is written. On TPC-DS, SSB, and NYC TLC (1,395 queries), intent signatures achieve an 82% hit rate versus 28% for text-based caching, with zero false hits.
There is also an extended version with more experiments and analysis.