Oracle

Info
Prompt
Loading Analysis 1
Analysis 1
Posts 1
Analysis 2
Posts 2

🌟 Inspiration

Our inspiration came from our experience as students regularly using ChatGPT. We recognized one of its limitations is in not understanding the latest culturally relevant contexts. We wondered whether integrating a large language model with the ability to search on social media could address this gap dynamically without relying on training with new data. This sparked our idea of a “Gemini Oracle" with encompassing knowledge of the human world through a more social lens—as long as that knowledge was being talked about online! Through Oracle's methodology, Gemini can receive selective glimpses into real-time discussions between millions of people worldwide without getting overwhelmed.

🤖 What it does

Oracle leverages Google Gemini's linguistic capabilities to analyze trending topics from a vast pool of ~500 million posts uploaded to X each day. It provides real-time assessment of public opinion and sentiment on current events and emerging cultural phenomena.

With Gemini, Oracle can generate key terms from news articles to search on X. It reads tweet timelines based on these searches and creates detailed sentiment analysis on user opinions around the subject. Gemini's access to real-time discussions from millions of users enables it to generate information about public opinion on the latest topics, such as natural disasters, events, and music.

🛠️ How we built it

We built our front end using React, JavaScript, and the MaterialUI library. We prioritized simplicity and efficiency, focusing on keeping the interface intuitive and easy to use. The project uses RapidAPI to access real-time data on news articles through NewsAPI and to search posts on X (formerly Twitter) through TwitterAPI. For generating search terms and analyzing X posts, we use the Gemini 1.0 Pro model to receive specific prompts that we concatenate with the extracted data as input and produce dynamic responses.

💻 Challenges we ran into

We encountered challenges with working within API call limits on RapidAPI, requiring us to manage our testing and app usage carefully to avoid exceeding the limits. Additionally, we faced issues with Gemini generating hallucinations, such as non-existent X posts or incorrect dates. To address this, we had to refine our prompts to emphasize certain instructions, such as "reference exactly the tweets from the data" or "BASED ON the data you analyze" for more accurate analyses. We also found that Gemini sometimes encounters errors with prompts containing certain sensitive search topics like violence and hate speech, or explicit phrases mentioned in the extracted posts due to safety violations.

😊 Accomplishments that we're proud of

We're proud to have successfully deployed our project with the functionalities we envisioned. We've been particularly fascinated by Gemini's ability to generate sentiment analyses on even the latest pop culture topics when equipped with posts from X. When compared to asking Google, ChatGPT, or Gemini alone, Oracle has shown more updated and accurate insights on numerous searches, specifically involving niche topics like celebrities and more recent events like the earthquake in New Jersey or the late April 8 solar eclipse.

💡 What we learned

Through working with Gemini, we've learned to refine prompts to better fit the pattern recognition of large-language models. For instance, we initially provided template examples for responses, but found that this often led to more hallucinations. We learned that concise and specific prompts were less distracting and guided Gemini best towards predicting a more contextually accurate continuation of the user-chatbot conversation. Additionally, we gained experience managing API in a React app, source control with GitHub, and hiding .env files within GitHub repositories. These tools are relatively new to us, and we enjoyed learning and improving while developing the project.

🚀 What's next for Oracle

In Oracle's future, we aim to integrate additional social media sources such as Instagram, TikTok, and Reddit, to offer more comprehensive and accurate analysis results. We also want to implement an efficient method to incorporate Gemini’s reading of images and videos as well. Before moving forward, we plan to further research the implications of a tool like Oracle, which we recognize could introduce biases, misinformation, and alter dynamics of social media and search. However, the potential to merge AI with social media and search on a larger scale is remarkable and can be an incredibly powerful tool, offering quick insights into collective public opinion that may be helpful or entertaining, even for the average person.