PlanLLM: Planning based Large Language Agents to Address Data Manipulation Queries, Using Local Large Language Models

Planning architecture are a powerful set of tools utilised to realise step wise execution of queries. These execution steps are often arranged in a topological order, with a fixed execution cycle. Generation and creation of a consistent set of plans for execution of a query, require several calls to Large Language engine, and can turn out to be highly uneconomical for industry use case. In this proposed system, a planning architecture that utilises local Large Language instances through Ollama client and Kubernetes pods to generate instruct model based responses, as well as a code generator and executor engine relying on coding based Large Language Models, including CodeGemma and CodeLlama are studied as an alternative to API based cloud solutions. This can increase the overall control of data flow for several use cases, as well as provide an exhaustive tool for data visualisation and manipulation tasks. This setup also provides a general architecture for using planning based agents for other such sequential use cases.