Application for review Whine For many years, he has provided helpful information to guests and other consumers. He experimented with machine learning from an early age. During a recent explosion in AI technology, he still encountered obstacles because it worked to use modern large language models to power some functions.
Yelp realized that customers, especially those that from time to time used the application, had a problem connecting to its AI functions, corresponding to AI powered assistant.
“One of the obvious lessons we have seen is that it is very easy to build something that looks cool, but it is very difficult to build something that looks cool and is very useful,” said Venturebeat Craig Saldanha, product director at Yelp.
Certainly not all the things was hard. After starting the Yelp assistant, his assistant to search for AI service powered, in April 2024 to a wider swath of customers, Yelp saw utility numbers for artificial intelligence tools that really begin to fall.
“The one who surprised us was when we launched it as beta for consumers – several users and people who are very familiar with the application – [and they] I loved it. We got such a strong signal that it would succeed, and then we swept it to everyone [and] The performance simply fell – said Saldanha. “It took us a very long time to determine why.”
It turned out that more free Yelp users, those that sometimes visited the website or application to find a recent tailor or plumbing, didn’t expect that they immediately talked to AI representative.
From easy to more involved AI functions
Most people know Yelp as a website and application for searching for restaurant reviews and menu photos. I exploit yelp to find photos of food in recent restaurants and see if others divide my feelings about a particularly bland dish. It is also a place that tells me if I plan to use a cafe as a working area for the day, it has wi -fi, plugins and seats, rare in Manhattan.
Saldanha remembered that Yelp invested in the artificial intelligence “for most of the decade.”
“I would say that on the 2013-2014 time line we were in a completely different generation of artificial intelligence, so we focused on building our own models to do such things as understanding the query. Part of the work consisting in establishing a significant relationship is helping people to improve their own search intentions, “he said.
But as AI evolutions, the needs of yelp. He invested in artificial intelligence to recognize food in photos sent by users to discover popular dishes, and then launched recent ways of connecting with traders and services and help in searching for users on the platform.
Yelp assistant helps Yelp users find the right “professional” for work. People can touch chatbox and use prompts or enter the task they need. Then the assistant asks questions more about the narrowing of potential service providers before developing messages to professionals who might want to bid for a job.
Saldanha said that professionals are encouraged to respond to users themselves, although he admits that larger brands often have telephone centers that support messages generated by the AI Yelp assistant.
In addition to the Yelp assistant, Yelp launched insight and the most vital information. LLM analyzes the mood of users and reviewers, which Yelp collects in the results of sentiments. Yelp uses a detailed GPT-4O line to generate a data set for the list of topics. Then it is refined with the GPT-4-Mini model.
The review function, which presents information from the review, also uses LLM prompt to generate a data set. However, it is based on GPT-4, with refinement with Turbo GPT-3.5. Yelp said he would update the GPT-4O and O1 function.
Yelp has joined many other LLM firms to improve the usefulness of reviews, adding higher search functions based on customer comments. For example, Amazon launched RufusAI powered assistant who helps people find advisable items.
Large models and performance needs
In the case of many recent AI functions, including AI assistant, Yelp turned to the Openai GPT-4O and other models, but Saldanha noticed that regardless of the model, Yelp data is a secret sauce for its assistants. Yelp didn’t want to locate himself in one model and retained the open mind, which LLM will provide the best service for its clients.
“We use OpenAI, anthropic and other models based on AWS,” said Saldanha.
Saldanha explained that Yelp has created a column to test the performance of models in correctness, meaning, awareness, customer safety and compliance. He said “these are really the best models” that worked best. The company runs a small pilot with each model before it takes into account the delay in costs and iteration.
Teaching users
Yelp also began joint efforts to educate each free and advanced users to feel comfortable with recent AI functions. Saldanha said that one of the first things they realized, especially at the AI assistant, is that the tone must have felt human. He couldn’t react too quickly or too slowly; It cannot be too encouraging or too rough.
“We put a lot of effort to help people feel comfortable, especially with the first answer. It took us almost four months to make the second piece. And as soon as we did it, it was very obvious and you could see that hockey sticks to the engagement, “Siedanha said.
Part of this process consisted in training the YELP assistant to use specific words and positive sound. After all this tuning, Saldanha said that they finally see higher numbers of use for AI Yelp.