RSS
Логотип
Баннер в шапке 1
Баннер в шапке 2
Project

"VkusVill" launched a bot based on artificial intelligence, which will select ready-made dishes

Customers: Izbenka (VkusVill)

Moscow; Trade

Contractors: AllSee
Product: YandexGPT (YaLM 2.0)
Second product: Postgres Pro DBMS

Project date: 2024/02  - 2024/03

2024: Launch of the helper bot "nIIhorn"

On April 2, 2024, VkusVill announced the launch of the PIIhorn bot-assistant, which, using artificial intelligence, helps customers select ready-made dishes based on their preferences.

source = TasteWill
nAIhorn

According to the company, the idea of ​ ​ creating a "PIE" was born as part of a brainstorm inside the creative team of VkusVilla during the development of special projects for the advertising campaign "Ready Food," aimed at highlighting the variety of cooking presented in VkusVilla. During the brainstorm, the team came to the conclusion that there are so many ready-made food positions in VkusVille (more than 1000, taking into account the "Burnt" assortment) that it is difficult for customers to figure out and choose what they want today. The solution was a bot, which, with the help of leading questions, helps the buyer choose a suitable dish for any meal.

The PIIhorn bot is designed on, database Postgres as well as a microservice architecture based on FastAPI and. Redis They allow the bot to operate stably through a scalable backend. The latest Aiogram 3 library is responsible for responsiveness, engagement and performance inside the bot, which makes it possible to asynchronously process requests. algorithm He is responsible for selection from, GPT-4 and OpenAI for conversational-understanding functions -. YandexGPT The bot itself works on dedicated servers ecosystems. Yandex Cloud The team helped with all the development and launch. AllSee

The development of the bot took a relatively short time: the start of the project was in mid-February 2024, and already on March 18, 2024, buyers got access to the PIrozhka. Everything related to bots and any products of neural networks is in constant revision, based on feedback from users, so this version of the assistant, strictly speaking, cannot be considered final.

File:Aquote1.png
The current implementation of the bot includes two parts: the first is the implementation of the bot-its responsiveness, high-load. And the second is intellectual functions, that is, selection of dishes and conversation with the user. The main task that faced us was to use the GPT model, OpenAi's GPT 4, to process the database with many "dry" categories and translate it into questions and answers that allow you to judge the preferences of buyers.

told Yegor Krasilnikov, CEO of AllSee.team
File:Aquote2.png

The bot's algorithm is as follows: it asks a limited number of questions to the buyer, and based on several answers, it is already ready to offer a number of dishes. At the same time, the buyer has the opportunity to continue communicating with the PIIhorn and deepen his answers. In this case, the selection of cooking will be more customized. The assistant sends from one to three dishes, and the user can go to the VkusVilla website or mobile application to place an order. You can also ask the bot to make another selection of dishes if you did not like the previous one. Inside the "PIIhorn" there is an opportunity to put a reminder the next day so that the bot reminds you to have breakfast or lunch.

File:Aquote1.png
We receive feedback from our buyers, and thanks to it we know how we can optimize the PIIhorn. Some are faced with the fact that the bot, like all neurons, can be biased. We monitor all requests, record and try to adjust.

told Elena Timofeeva, Communications Manager marketing of VkusVill
File:Aquote2.png