Recently, LlamaIndex introduced a new feature called <a target="_blank" class="af pa" href="https://docs.llamaindex.ai/en/stable/understanding/workflows/” rel=”noopener ugc nofollow” target=”_blank”>Workflow in one of its versions, which provides event-based and logical decoupling capabilities for LLM applications.
In today's article, we will delve into this feature through a practical mini-project, exploring what's new and lacking. Let's get started.
More and more LLM applications are shifting toward intelligent agent architectures, expecting LLMs to satisfy user requests through calls to different APIs or multiple iterative calls.
However, this change brings with it a problem: as agent applications make more API calls, program responses slow down and code logic becomes more complex.
A typical example is <a target="_blank" class="af pa" href="https://docs.llamaindex.ai/en/stable/api_reference/agent/react/#llama_index.core.agent.react.ReActAgent” rel=”noopener ugc nofollow” target=”_blank”>ReactAgentwhich involves steps like Thinking, Action, Observation and Final Response, which require at least three LLM calls and one tool call. If loops are needed, there will be even more I/O calls.
