HATEOAS 与大语言模型结合应用
1 分•作者: charlieflowers•6 个月前
我只是观察到，灵光一现，想分享一下。<p>我所见过的说自己在“做 REST”的大部分开发团队，实际上并没有遵循 HATEOAS。按照 Roy Fielding 的严格定义，他会认为这“算不上真正的 REST”。（现在别分心，我不想卷入那场纯粹主义的争论）。<p>许多人没有做 HATEOAS 的原因是，它要求 API 客户端具有智能和适应性。它需要发现“好的，接下来我可以做什么”，应用逻辑，然后选择下一步。但许多公司的时间都很紧，把 REST 简单地理解为“通过 HTTP 传输 JSON，并使用一致的 URL 模式”要简单得多。<p>酷的是：有了 LLM 的加入，HATEOAS 就被解放了。LLM 可以做到“愚蠢”的 API 客户端做不到的事情：询问“接下来我可以做什么”，然后使用_推理_来理解这些选项并选择一个。
查看原文
Just an observation, a light bulb moment, I wanted to share.<p>Most of the dev teams I've ever encountered who said they were "doing REST" were not actually following HATEOAS. Per a strict reading of Roy Fielding, he would consider that "not really REST." (Now don't get distracted, I don't want to wade into that whole purist debate).<p>The reason many did not do HATEOAS is that it requires the API client to be smart and adaptive. It would discover "ok, what can i do next", apply logic to it, and choose the next step. But many shops were on tight time commitments and it was much simpler to just think of REST as "json over http with consistent url patterns."<p>The cool thing is: With an LLM in the mix, HATEOAS is unchained. An LLM can do exactly what a "dumb" api client cannot: ask "what can i do next", and then use _inference_ to understand those options and select one.