case-study - 2023-06-21
Case Study: How Lastro built a chatbot with Metal
by Allan Paladino & Taylor Lowe
Lastro x Metal
Lastro is a real estate software company that simplifies workflows in buying and managing properties. As an all-in-one platform, it’s crucial that they provide buyers, renters and property owners with up-to-date information as easily as possible. This is no small feat, as there are over 5 million houses in their home city of São Paulo alone!
With all of the information available in their knowledge base and the advent of LLM-powered chatbots, they knew there was an opportunity to give their users a better experience.
Testing off-the-shelf and homegrown tools
To validate their thinking, Lastro wanted to get something in front of users as quickly as possible.
They first tried an off-the-shelf chatbot tool. While this made it easy to get an initial version out the door, they quickly found themselves limited by the lack of flexibility and specific features their users needed. In particular, it was difficult to understand how their embeddings were being generated and retrieved. Query results were just a little off, and it was hard to understand why.
Being a strong technical team, they next tried to build a knowledge base retrieval solution themselves. The team built a vector database, generated embeddings, and layered on semantic search capabilities to simulate user queries (did I mention they’re a strong technical team?). The results they got back were a little better, but still not quite good enough. At this point, they realized that this was a problem other people may already be working on, so they started looking for solutions and found Metal.
Building for production with Metal
It was easy for Lastro to get started with Metal, as embedding generation, chunking techniques, storage, and retrieval all come out-of-the-box. Using Metal’s retrieval functionality, the team was able to select the most relevant information given a query and then send that data to a GPT model to provide the chat experience.
The first thing they noticed was the quality of the results. Where other solutions would often return data that was irrelevant to a query, Metal consistently delivered accurate matches for their dataset. Queries about real estate prices were especially tricky with the off-the-shelf and homegrown solution, where Metal returned the correct results.
Furthermore, Metal’s developer experience and UI made it easy to quickly test different index methods and confirm if queries were executing correctly. Even non-technical team members could understand how datasets were being stored and retrieved. This allowed their product team to quickly build and ship chatbots, improving the experience for their users with each iteration.
Introducing Lais, powered by Metal
Their chatbot was now ready for production and was released with the name Lais. Despite very little promotion, they were floored by the rapid adoption. Lais gained 2,000 users in the first week alone, and after a month of usage, that number has grown to over 5,000 and counting! Lastro was even contacted by major real estate companies in the area, asking to learn more about Lais and how they could use it.
Screenshot of the Lais.ai chatbot
What was more important to the Lastro team, however, was seeing how their users interacted with the product. They noticed many users were chatting with Lais as if it were a real estate agent, asking questions about properties to buy and rent – even though Lais was not designed for this.
Mobile screenshot of the Lais.ai chatbot
Observability into their user’s behaviors was crucial to inform Lastro which direction they wanted to take the product next.
Future plans for Lais
After so much early success, Lastro realized that this was a huge opportunity to create the best personal assistant for tenants and buyers to find their next property, turning Lais into an AI-powered personal assistant specialized in real estate.
Metal helped Lastro overcome the limitations of traditional data storage, which omit the semantic and subjective meaning of these queries – a key benefit of using embeddings and a vector database in a solution like Metal. The team at Lastro plans to fully take advantage of Metal’s ever increasing query abilities, providing new and exciting experiences for their users.
If you’re interested in learning more or building a similar application, please get in touch with us. We would love to help you build!