Assignment 5 Implementing a Simple Retrieval-Augmented Generation (RAG) System

By Evelyn Bushell

Models Used

LLM Model

google/flan-t5-small

Embedding Model

sentence-transformers/all-MiniLM-L6-v2

Retrieval Process

The Knowledge Base is split into 5 chunks each with 300 characters. The retriever then compares the query with each chunk and grabs the chunks with the highest relevancy scores and adds them to the prompt.

Results Analysis

This does seem to be very fact based. The model is extremely hesitant to provide an inaccurate answer and generally sticks to the context given to it, while still able to make obvious observations such as the chosen film being more sad than happy.