Abstract
Using large language models (LLMs) to query large, scale, graph, structured data can present challenges, as encoding an entire graph into a single prompt may exceed context window limits and can lead to reduced accuracy on multi-hop queries. Systems and methods can partition a graph into a collection of smaller, human-readable documents, or shards. An agentic orchestration process may use a global symbolic index, contained within a summary shard, to locate and retrieve relevant shards on demand. This process can decompose a complex query into a sequence of single-hop LLM calls, where each call can operate on a context comprising one or more shards. This approach can enable an LLM to function as a query engine for graphs that are substantially larger than its context capacity, which can improve scalability and may mitigate a need for a dedicated graph database system for certain analytical workloads.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kommuri, Sai Charan Tej, "Agentic Graph Querying on a Sharded Architecture Using a Large Language Model", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10361