Together AI Showcases LLM-Based Approach to Database Query Optimization

According to a recent LinkedIn post from Together AI, the company’s research unit is exploring how large language models can correct suboptimal database query plans to improve execution speed. The post describes a benchmark framework, DBPlanBench, that exposes a database’s physical operator graph to an LLM, which then applies targeted JSON patches to adjust join ordering rather than regenerating full plans.

The LinkedIn post highlights reported performance gains on standard TPC-H and TPC-DS workloads, including up to 4.78x speedups on complex multi-join queries and improvements of more than 5% on 60.8% of queries tested. It also notes significant reductions in build memory usage on specific benchmarks and indicates that optimized plans at small scale may transfer to larger databases, with both paper and code released as open source.

For investors, the post suggests Together AI is positioning its research at the intersection of generative AI and database optimization, an area with potential relevance for analytics-heavy enterprises and cloud data platforms. If this approach proves robust and generalizable, it could expand the company’s addressable market into performance tooling for data infrastructure, potentially enabling product offerings or partnerships focused on reducing compute costs.

The emphasis on open-source code and research may also reinforce Together AI’s visibility within the developer and academic communities, which can be a strategic channel for adoption of its broader AI platform. However, the LinkedIn content does not provide commercialization timelines, revenue models, or customer traction, so the direct financial impact remains uncertain and will depend on future productization and integration into real-world database environments.