Inspiration
The inspiration for MaPT stemmed from the observation of the abundance of Map/GIS data available online. Our team envisioned leveraging this wealth of information to create a unique interactive experience by leveraging artificial intelligence and LLMs. The recent release of the Assistants API from OpenAI sparked their curiosity, leading to combining conversational AI's power with mapping data. The goal was to make GIS data more accessible by allowing users to interact with the map in a natural and conversational manner.
What it does
MaPT redefines the online mapping experience, akin to platforms like Google Maps. It introduces a unique twist: users engage with a conversational chatbot rather than conducting direct searches. This intelligent chatbot not only responds to user requests but also marks locations on the map. MaPT stands out by effortlessly tackling complex questions, providing intricate details about buildings, and maintaining a fluid conversation with users, retaining context between queries.
Consider this scenario: a user poses a question such as "What is the largest building in Marietta, GA?" MaPT swiftly identifies the building and marks its location on the map. Following this, the user seamlessly inquires, "Is there a park within 3000 meters of this building?" MaPT not only undertakes the task of locating the park but also marks its position on the map. This interactive process showcases MaPT's ability to handle consecutive queries, providing users with a smooth and informative experience.
How we built it
Our web app's front end is crafted with React and features an interactive map powered by Leaflet.js. This seamlessly connects to a robust Flask backend, facilitating communication with an OpenAI assistant. The assistant, equipped with a GIS database and a tailored prompt for user queries, efficiently processes questions. It then returns not only the answers but also a set of annotated coordinates to the front end, ensuring a smooth and responsive user experience.
Challenges we ran into
Navigating diverse data sources posed a significant hurdle in our project. The online data we encountered often provided expansive information at the county or state level, falling short of the granularity required for individual buildings—the focal point of our endeavor. After curating a substantial dataset, our next challenge lay in meticulous data cleaning. This involved refining the dataset to ensure seamless comprehension by our Large Language Model, a crucial step in enhancing the project's overall effectiveness.
As the project progressed, we encountered a notable challenge related to the rigidity of the GPT model in interpreting names. A specific issue arose when the model distinguished between variations such as "Fort St" and "Fort Street," resulting in failed queries and a lack of corresponding entries. This discrepancy highlighted the need for a nuanced approach to accommodate such variations in naming conventions.
To address this challenge, the team implemented a solution by refining and tuning the prompt used to interact with the GPT model. This involved introducing adjustments that enhanced the model's flexibility and recognition capabilities, allowing it to understand and respond effectively to queries despite variations in street name formats. The iterative process of refining the prompt proved instrumental in overcoming the naming discrepancy, ultimately leading to improved query success rates and a more accurate interpretation of user input. This experience underscored the importance of fine-tuning the interactions with language models to ensure adaptability to diverse naming conventions and variations, contributing to the overall success of the project.
Accomplishments that we're proud of
We were genuinely impressed by MaPT's adeptness in tackling challenging queries, particularly its proficiency in identifying street intersections. Its robustness exceeded our initial expectations, showcasing an impressive ability to discern user intent even when faced with misspelled or abbreviated building names. A noteworthy achievement in our project was effectively leveraging the functions feature of the OpenAI Assistants API. This allowed us to skillfully process the output generated by the OpenAI models, enabling our web app to receive a meticulously formatted list of points for annotation. This strategic use of API capabilities significantly enhanced the functionality and efficiency of MaPT, contributing to its overall success in providing accurate and nuanced responses to user inquiries.
What we learned
Our experience with the OpenAI Assistants API was enlightening, revealing its vast potential for innovative applications. Delving into the intricacies of interactive maps on the front end provided valuable insights, as did navigating the conversion process to and from GIS data. One notable challenge involved refining user input for seamless conversion into GIS data—a process that proved to be intricately nuanced and highly educational. This journey of exploration not only deepened our understanding of the OpenAI Assistants API but also enriched our knowledge of the complexities inherent in interactive mapping and data conversion.
What's next for MaPT
Presently, our project operates on a constrained dataset specific to the city of Marietta, GA. To enhance the capabilities of our assistant and broaden its utility, our immediate goal is to integrate a more extensive GIS dataset. This expansion is pivotal as it will empower MaPT to respond to a wider array of user queries, encompassing diverse geographical aspects beyond the current scope.
Looking ahead, we aspire to enrich the user experience by introducing additional features for interacting with the map. A key development in our roadmap involves implementing the functionality for users to selectively choose individual buildings on the map. This enhancement will allow users to pose specific inquiries or seek detailed information about particular structures, further personalizing their engagement with MaPT.
By incorporating more comprehensive GIS data and refining the user interaction options, we aim to transform MaPT into an even more versatile and user-friendly tool. These planned advancements align with our vision of providing users with a robust mapping platform that caters to a broader spectrum of inquiries and offers a tailored and interactive exploration of geographical data.
Log in or sign up for Devpost to join the conversation.