Enhanced Medical RAG Bot

• 

  Main page

• 

  image analysis page

• 

• 

• 

• 

Inspiration

As international students, we have all encountered the frustrating reality of navigating the healthcare system. Getting a doctor's appointment can take months, and for many of us, even basic medical queries are met with uncertainty. Often, it’s difficult to know who to refer to or where to get answers quickly. This gap in accessibility to reliable and immediate healthcare information is what inspired us to build a solution:

What it does

This application allows users to ask day-to-day medical questions, and the bot provides responses based on past doctor-patient communication data. Best part the dataset is near real time. it keeps on updating on daily basis. By mimicking real-life doctor-patient exchanges, we aim to provide quick, accurate, and relevant answers to common healthcare questions, saving users both time and frustration.

How we built it

We have developed a system that combines several components to offer real-time, accurate responses to user queries. The frontend is built using Streamlit, which provides an interactive user interface. For the dataset, we are leveraging a near real-time data stream sourced from Hugging Face. This dataset is regularly updated to ensure that the information remains current and relevant.

When a user submits a search query, the system retrieves and compares it against a database of pre-existing questions. To enhance the accuracy of this comparison, we use Facebook's search functionality, which helps in identifying relevant content from the broader web. The results from the Facebook search are then processed and passed into a large language model (LLM), which is responsible for generating well-formed, informative answers based on the search results.

This approach ensures that user queries are met with the most relevant, up-to-date information, while also leveraging the power of AI to craft coherent and contextually appropriate responses.

Challenges we ran into

• finding closest searching algorithm and indexing data
• mongodb collection setup with streamlit
• performing image analysis

Accomplishments that we're proud of

• we have accuracy of 90% in answering the right answers to user queries with high speed. provide additional information to user queries

What we learned

• LLMs, image analysis, sentence transformers, vector searches, FAISS

Built With

• azure
• faiss
• huggingface
• langchain
• llm
• microsoft
• rag