请问HN：大型语言模型的提示工程
我正在进行一个项目，需要提取用户意图，并将其转移到确定性的工具/功能/API 执行中，随后通过另一组工具对结果进行细化和转换。由于收集正确的意图和参数（潜在提示中存在许多微妙的差异）相当具有挑战性，因此我使用了一系列连续执行的提示来微调，以准确收集所需的信息，以便进行相对可靠的工具执行。我尝试过使用多个代理框架（包括 langchain/langgraph），但情况很快变得混乱，这种混乱容易导致许多副作用。
因此，我想知道是否有工具、方法或其他任何东西可以更好地控制 LLM 执行链，以避免陷入混乱的配置和/或代码执行实现中？也许还有更直观的解决方案，还是我只是一个人在为此苦恼？
查看原文
I’m working on a project where I need to extract user intents and move them to deterministic tool/function/api executions + afterwards refining/transforming the results by another set of tools. Since gathering the right intent and parameters (there are a lot of subtle differences in potential prompts) is quite challenging I’m using a long consecutive executed list of prompts to fine tune to gather exactly the right pieces of information needed to have somewhat reliable tool executions. I tried this with a bunch of agent frameworks (including langchain/langgraph) but it gets very messy very quickly and this messiness is creating a lot of side effects easily.<p>So I wonder if there is a tool, approach, anything to keep better control of chains of LLM executions which don’t end up in a messy configuration and/or code execution implementation? Maybe even something more visual, or am I the only struggling with this?