Join our daily and weekly newsletters for the latest updates and the exclusive content on AI coverage. Learn more
Examination Jap Has provided useful information to guests and other consumers for decades. He had experienced automatic learning for his early years. During the recent explosion of AI technology, she still encountered stumbling blocks while working to use modern language models to supply certain features.
Yelp realized that customers, especially those who only used the application occasionally, had problems Connection with its IA featuresLike his assistant powered by AI.
“One of the obvious lessons we have seen is that it is very easy to build something that looks cool, but very difficult to build something that looks cool and is very useful,” Craig Saldanha, product manager at Yelp, in Venturebeat in an interview.
It was certainly not easy. After launching the Yelp assistant, his service assistant fueled in AI, in April 2024 to a wider band of customers, Yelp saw use figures for its AI tools that start to decline.
“The one who took us by surprise was when we launched this as a beta to consumers – some users and people who know the application very well – [and they] I loved it. We obtained a signal so strong that it would succeed, then we deployed it to everyone, [and] The performance has just fallen, ”said Saldanha. “It took us a long time to understand why.”
It turned out that the more occasional users of Yelp, those who sometimes visited the site or the application to find a new tailor or a new plumber, did not expect to speak immediately with an AI representative.
Simple features to be more involved
Most people know Yelp as a website and an application to search for restaurant criticism and menu photos. I use Yelp to find food photos in new restaurants and to see if others share my feelings on a particularly bland dish. It is also a place that tells me if a coffee that I plan to use as a workspace for the day has wifi, traffic jams and seats, a rarity in Manhattan.
Saldanha recalled that Yelp had invested in AI “for a better part of the decade”.
“Back, when I would say that in the 2013-2014 calendar, we were in a very different generation of AI, so our concentrating was to build our own models to do things like understanding requests. Part of the work to establish a significant connection is to help people refine their own research intention, “he said.
But as AI continued to evolve, Yelp’s needs too. He invested in AI to recognize food in images submitted by users to identify popular dishes, then he launched new ways of connecting to business and services and services and services Helps guide user searches on the platform.
The Yelp assistant helps Yelp users find the right “pro” with which to work with. People can press the chatbox and use the prompts or type the task they need. The assistant then asks for follow -up questions to reduce potential service providers before writing a message to pros who may want to bid for work.
Saldanha said the pros are encouraged to respond to users themselves, although he recognizes that major brands often have call centers that manage the messages generated by the Yelp AI assistant.
In addition to Yelp Assistant, Yelp has launched information and protruding facts from the journal. The LLMs analyze the feeling of users and examiners, which Yelp collects in feeling scores. Yelp uses a GPT-4O prompt To generate a set of data for a list of subjects. Then it is refined with a GPT-4-MINI model.
The revision highlights functionality, which presents information from reviews, also uses an LLM prompt to generate a set of data. However, it is based on GPT-4, with a fine adjustment of the GPT-3.5 Turbo. Yelp said he will update the functionality with GPT-4O and O1.
Yelp has joined many other companies Use of LLMS to improve the usefulness Notices by adding better research functions according to customer comments. For example, Amazon launched RufusAn assistant powered by AI who helps people find recommended items.
Large models and performance needs
For many of its new AI features, including the AI assistant, Yelp turned to the Openai GPT-4O and other models, but Saldanha noted that no matter the model, Yelp data is the secret sauce for its assistants. Yelp did not want to shut himself up in a model and kept an open mind on the LLMS would provide the best service to his customers.
“We use OPENAI, Anthropic models and other models on the AWS basis,” said Saldanha.
Saldanha explained that Yelp had created a section to test the performance of correction, relevance, consciousness, customer and compliance models. He said that “it is really the high -end models” that worked best. The company manages a small pilot with each model before taking into account the cost of iteration and the response latency.
Teach users
Yelp also embarked on a concerted effort to educate occasional and powerful users to make themselves comfortable with the new features of AI. Saldanha said that one of the first things they have done, in particular with the AI assistant, is that the tone should feel human. He couldn’t answer too quickly or too slowly; He couldn’t be too encouraging or too abrupt.
“We have put a lot of efforts to help people feel comfortable, especially with this first answer. It took us almost four months to do this second song well. And as soon as we did, it was very obvious and you could see that hockey will be held in the commitment, “said Saldanha.
Part of this process consisted in forming the assistant Yelp in using certain words and appearing positive. After all this fine setting, Saldanha said they finally saw higher use numbers for Yelp’s AI features.
