Highlights

• Increased the precision and quality of contract query answers by building a Contract Q&A RAG system using LangChain and OpenAI.

• Measured the RAG’s performance using metrics such as context precision, context recall, answer relevancy, and faithfulness from RAGAS.

• Improved the performance metrics of the RAG by implementing a data chunking strategy with RecursiveCharacterSplitter. Found that 500 chunk sizes had the best results over 1000 and 128 chunks.

• Tested LangChain’s advanced retrieval methods, including Multiquery, Ensemble, and ParentDocument Retriever, to optimize performance and ultimately built the RAG with the Multiquery Retriever which had the best performance. 

• Developed a user-friendly interface with Streamlit, allowing users to efficiently interact with the RAG system.

• Fine-tuned the Llama2 13B quantized model for Swahili text classification in 5 categories, including sports, entertainment, business, local news, international news, and health. As a result, the project contributes to better language processing for Swahili, an underrepresented language. 

• Collected and combined 54k rows of Swahili news data through web scraping and from Hugging Face and African NLP datasets. 

• Evaluated model performance on text generation, classification, summarization, question answering, and translation before fine-tuning. 

• Preprocessed the data and utilized SFTTrainer for supervised fine-tuning, pushing the final model to Hugging Face. 

• Leveraged individual strengths by working in a team of 5. My role involved data collection, preprocessing, and performing inference on the Llama2 13B quantized model.