feday2025-通过自学习用户行为让语音应用越用越聪明.pdf

1、LoisToma|19/12/2025通过学习户为，让语应越越聪明Building 10 x smarter voice AI application by self-learning through realtime conversation我有个问题Lois|赵恩The bundler girl works on bundler thingOr companies with intense bundler problemsCurrently Toma,prevZephyr Cloud,Daytona Module Federation|React Native Runtime OTA|Me

2、tro Bundler|Repack Automotive industry|Million conversations over PSTNChasing time with voiceLLMTelephony ProviderIntegrationRedisPostgresEarliest function call find_availability:You are required to run find_availability when users asks for bookings.,book_appointment:Once user confirm a time slot yo

3、u should book an appointment.”中企业客户都需要定制化服务100%0 operation leftWe fired the entire support teamHow can we build a system that able to learn from our users?How can we build a system that we are able to penetrate every industry?早期实现const msg=db.interactionMessages.findMany(type:in:human,humanAssistant

4、 )const output=new LLM.askStructuredQuestions(model:MODELSTHINKING,question:Here are some chat messages our customers have,can you help us generate some policies?$msg.join(,)有个问题有个问题-Token Overflow，上千万聊天记录每条都要通过Evaluation-过于细节化-Large amount of duplication and conflicts-学习到的内容和客户的需要不样-捷径，没有合理的评估标准-Ru

5、ns in Temporal Workflow every 3 minutes,天时间花费上万美不符合标准的Policy样本-When customers asks to replace windows and provide a quote,the price is$500.-When no booking slot is available,ask customers to come in and drop the car.-2023 Nissan Rogue recall R25A8(engine bearing)begins with a software reprogram;furt

6、her work only if a specific code sets.Embedding search doesnt work that wellwith Postgres处理不合规的PolicyCall ClusteringEvery minutes Database Trigger Event-Embedding Creation Both conversations and PoliciesEvery Day At 8pm PST Call batching-k-