Project:

• TAMU Datathon 2024 first place winner Connections AI model.
• Grouper for solving New York Times Connections using contextual and semantic similarity.
• Flask Application for Category Detection, Single-Word Prediction, and Multi-Word Predictions with 6 fallback algorithms.

Description:

• Developed a Flask API that leverages OpenAI and spaCy NLP to process word sets, identify matching categories, and suggest solutions for word puzzles based on semantic similarity and wordplay. The application enables high-performance matching and word prediction through caching and optimized vector operations.
• Also checks for possible anagrams, palindromes, and word puzzle similarity.

Key Features:

• Category Matching: Matches user-submitted words to pre-defined categories, using exact matching, semantic similarity, and heuristics.
• Word Prediction: Utilizes cosine similarity on word vectors to predict missing words in a set and suggest potential category memberships.
• OpenAI Integration: Employs OpenAI's GPT model for advanced word puzzle hints, especially for anagram and semantic challenges.
• Data Processing: Reads category data from JSON files, standardizes word format, and uses word vectors to compute contextual similarities.
• Performance Optimization: Implements caching with lru_cache to reduce repetitive processing and improve response times for frequently used NLP operations.
• Error Handling & Logging: Includes robust error logging for data loading and API errors, ensuring reliable performance.
• Ensures the same answer isn't returned twice on the rare occasion of an incorrect guess.
• Given 3 words and a "One Away" hint, the fourth word is predicted using semantic similarity. If less than 3 words are correct, it falls back to GPT4o; if it is still incorrect, it falls back to a third model.

Technologies Used:

• Backend: Python, Flask, spaCy (en_core_web_md model), OpenAI API, NumPy
• API: JSON-based RESTful endpoints for client interaction
• Utilities: Logging for performance monitoring and troubleshooting

Example inputs and outputs:

• A list of remaining words, correct guesses, incorrect guesses and whether the previous guess was OneAway from the true answer is passed to the API.
• The API is designed to return a list of four elements as the next best guess. The reason for designing this project using an API is because the datasets were large and we wanted to mask our API keys.

Usage:

• Clone the repository and run apilauncher.py to start the API. Pass in command line arguments for the correct dataset and expanded dataset using python apilauncher.py
• 2 example datasets have been provided containing categories used in previous games and an expanded data set to predict future games.
• Ensure that app.py has the correct IP address given present in the function calling the API and provide your OPENAI api key in apilauncher.py.
• Run app.py and then run evaluator.py to evaluate the default test cases.
• Other hidden randomized test cases were passed into the function to conclude our project as the winner.

Potential improvements for the future:

• isOneAway is typically hit during a failed test case. Creating new logic providing context for isOneAway and categories to a predefined LLM (for cases where 3 are not in a dataset) will improve the accuracy significantly (however, we did not implement this as the deadline for the datathon had finished).
• Expanding the larger dataset is key for fast performance and more accurate results. We only expanded around half the categories for this project due to time constraints.