This project implements a Retrieval-Augmented Generation (RAG) pipeline using LlamaIndex, OpenAI's GPT-3.5, and FAISS, designed to allow real-time, grounded Q&A from uploaded documents (PDFs, TXT, DOCX). It runs entirely in Google Colab, with a future-ready path to a UI (Gradio, Streamlit, etc.).
- Ingests your documents (e.g. resumes, legal docs, manuals)
- Parses and chunks them using
SimpleDirectoryReader - Converts text chunks to vector embeddings using
OpenAIEmbedding - Stores the embeddings in FAISS for high-speed retrieval
- Retrieves top-k relevant chunks per user query
- Feeds them into
gpt-3.5-turboto generate grounded answers - Displays the response along with source text snippets
| Layer | Component | Description |
|---|---|---|
| Document Parsing | SimpleDirectoryReader |
Automatically extracts and chunks text from .pdf, .txt, .docx files |
| Embedding | OpenAIEmbedding |
Converts each chunk into a high-dimensional vector |
| Vector Store | FAISS |
Stores all embeddings in-memory and enables top-k similarity retrieval |
| Index Construction | VectorStoreIndex (LlamaIndex) |
Maps document chunks → FAISS and connects them to the LLM orchestration |
| Query Execution | GPT-3.5 Turbo |
Retrieves top-k chunks, constructs a prompt, and generates a grounded response |
Colab link: https://colab.research.google.com/drive/1HVmdv9OueWyUQdvS1D5H-Fqh3Yc6Zpcy?usp=sharing