The project is RAG
Task: We have a technical support department in our company, and we needed to make a RAG for it, which would help the technical support department to solve different problems with 1C faster.
Result: As LLM we tried 2 options - training ru-gpt3.5 and gigachat-pro and finally we decided on gigachat-pro as it showed the best results. We used QDrant and Elasticsearch. To get embeddings we used our internal model which was already well-trained on our data and gave good embeddings. in the end we got MAP@3 = 0.48.
Technologies: Python, Torch, Pandas, Numpy, Machine Learning, Transformers, CatBoos,t XGBoost, Seaborn, Docker, Airflow,DeepSpeed, LangChain