AI Start: Building a Search Function
By Nico Lutz
AI Start is a joint initiative between Katapult, Seedstars and Bakken & Bæck, with the goal of exploring the potential of AI in providing detailed analysis of early-stage start-up investments. The project seeks to encourage a more comprehensive outlook by considering a range of factors that impact the success of start-ups.
A major goal was to develop an AI engine by integrating qualitative and quantitative data from newly developed features in order to predict the likelihood of raising a Series A/B investment round. By acknowledging that every start-up undergoes a distinct journey, and that achieving success is shaped by numerous factors beyond financial aspects, we focused on BB's strong design and engineering background to rethink the process of conducting searches with the aim of finding suitable start-ups in a pool of applicants. Here we present some of our findings and showcase our results.
Search
What is a search? Put plainly, it's the process of typing a question into a search box and then showing the results. But the advent of LLMs offers new possibilities and gives users a new perspective on the search function.
After initial experiments we found that LLMs are indeed capable of transforming natural language queries into plain structured query language (SQL), for example, which in turn could be used to query the database directly. Specifically, the question is used in a prompt template with background information about the underlying data that allows the LLM to perform rich queries to the database in order to retrieve the desired information. The user now has full control over what is queried to the database. But this also leads to problems related to SQL and prompt injection that need to be considered. More importantly, however, the user can ask questions that result in nonsense queries or that repeatedly return zero results. Our experiments show that it's important to guide the user in these cases and to show immediate feedback, stating the reason for the problem, so as not to frustrate the user.
We also explored the possibility of using the LLM to retrieve suggestions based on previously asked questions and the underlying data structure — this can give users hints and prompt them to search in another direction. Keywords can also immediately be transferred to the correct question by using the previously used context.
Localising start-ups by asking the system for information with natural language questions.
Dashboard
The dashboard is the heart of the application we built. We focused on both clarity and user-friendliness; it's designed to present a wide range of data points in a visually simplified format. Users have multiple options in order to find a suitable candidate. Searching in a pool of data points is a common task for anyone working with data, but our goal was to rethink how searches can be made by leveraging large language models (LLMs) and tools from data visualisations.
Visualisations
Another approach to searching is the use of data visualisation techniques to give the user a direct overview of a cluster of points, for example. We specifically used maps and scatter plots with different axes to show the wide variety of options. Pre-selecting a certain feature, such as a Sustainable Development Goal score, then mapping this against another feature can help users find the best company for their purposes — for example, with a high SDG Impact Rating and high revenue. Colour-coding the results can give a clear overview and insights into the data. More fine-grained control can be gained by selecting sub-sections and further sub-selecting features.
While text searches and visualisations can take us a long way, we also introduced a more traditional search box where users can (de)select filters and then select companies in a list view.
Advanced search functions can be paired with traditional search boxes.
Heat map with variable x and y axes. Here the user can select a sub-section for further detailed inspection.
Maps
Since we intend to use AI Start globally we had to be able to make sense of feature values in the context of a company's geographic region. A map helps users locate data points across the world. Selecting certain features can also help them navigate the landscape of companies across a certain region. Features like revenue or income must be put into the context of a country or region and can't be compared across continents, for example.
Sub-selecting companies based on geography on a world map.
Rethinking our approach to search was both challenging and rewarding; it pushed us to leverage state-of-the-art technology to investigate the process. Exploring various visualisation techniques was an exciting part of our journey, enabling us to redefine how users interact with search results. This innovative approach not only enhanced the user experience but also opened up new possibilities for information retrieval. In the end, our efforts transformed the way people access information, making it a highly gratifying experience.