The objective of this research is to explore the use of Retrieval-Augmented Generation (RAG) in creating a personal chat-bot powered by pretrained large language models (LLMs). By utilizing Langchain as a vector database for efficient data retrieval, this study aims to demonstrate the practicality of open-source LLMs for personal, privacy-preserving use. The research will focus on minimizing computational resources, enabling users to interact with their personal data securely on their local devices.
Book - Encyclopedia of Foods: A Guide to Healthy Nutrition.pdf
| Embedding Models | all-MiniLM-L6-v2 | stella-base-en-v2 |
|---|---|---|
| Size | 0.08GB | 0.2GB |
| Embedding Dimension | 384 | 768 |
| Parameter | 22.7M | 55M |
| Max Tokens | 512 | 512 |
| Language | English | English |
| Retrieval Accuracy | 73% | 85% |
Embedding model with higher embedding dimension has high retrieval accuracy.
| Chat Models | Google/T5-Base |
|---|---|
| Size | 0.9GB |
| Parameter | 223M |
| Year | 2020 |
Langchain - FAISS
| Apple Silicon | M1 |
|---|---|
| TFLOPS | 2.60 |
Embedding models: Hugging face Leaderboard
Embeddings: Theory
Embedding Visualization: Visual Representation