一种可扩展的LLM方法，通过用户生成内容增强聊天机器人的知识

DoorDash’s support chatbot handles a huge volume of questions from Dashers and customers every day. Chats can range from guiding a Dasher to their next delivery and reassuring a customer about what’s happening when an order runs late to explaining new features as they launch.

DoorDash的支持聊天机器人每天处理大量来自Dashers和客户的问题。聊天内容可以从指导Dasher进行下一次交付，到在订单延迟时安抚客户，再到解释新功能的推出。

But as our marketplace grows, so does the complexity of these conversations. New policies, product changes, and a long tail of edge cases all demand fresh answers. Manually maintaining the knowledge base cannot effectively scale and is too resource-intensive and time-consuming.

但是随着我们的市场不断增长，这些对话的复杂性也在增加。新的政策、产品变化以及大量边缘案例都需要新的答案。手动维护知识库无法有效扩展，并且资源密集且耗时。

We needed a smarter solution. By pairing clustering algorithms with large language models (LLMs), we can surface the highest‑ROI content gaps automatically and draft accurate articles in minutes instead of weeks based on user-generated content, or UGC. This allows our team to focus on refining and elevating new content, while the heavy lifting of identifying gaps and drafting new material happens at machine speed.

我们需要一个更智能的解决方案。通过将聚类算法与大型语言模型（LLMs）结合，我们可以自动呈现最高投资回报率的内容空白，并根据用户生成内容（UGC）在几分钟内草拟准确的文章，而不是几周。这使我们的团队能够专注于完善和提升新内容，而识别空白和草拟新材料的繁重工作则以机器速度进行。

In this post, we walk through the system we built, the lessons we learned, and the impact we’re already seeing.

在这篇文章中，我们将介绍我们构建的系统、我们学到的经验教训以及我们已经看到的影响。

We begin by feeding thousands of anonymized chat transcripts into a semantic clustering pipeline, selecting only those conversations that were escalated to a live agent so that we can zero in on the cases where our chatbot fell short. The clusters that emerge highlight the issues causing the most friction for Dashers and customers, allowing us to rank gaps in the knowledge base, or KB, by both frequency and severity.

我们首先将数千个匿名聊天转录输入到语义聚类管道中，仅选择那些升级到现场代理的对话，以便我们能够聚焦于聊天机器人表现不佳的案例。出现的聚类突显了导致Dashers和客户摩擦的主要问题，使我们能够根据频率和严重性对知识库（KB）中的空白进行排名。

Figure 1. Escalated chat transcripts are automatically grouped into meaningfu...