This repository implements a local RAG-based AI bot for Upwork API documentation. It uses Streamlit for the interface, FAISS for vector storage, and sentence-transformers for embeddings. The bot reads PDFs, processes text, and uses the DeepInfra-hosted Meta-Llama-3.1-8B model to generate responses. Challenges included handling multi-column PDFs and managing chunk boundaries in code snippets.
From the README
Technical Summary — Upwork API RAG Support Bot --- Project Overview A fully local Retrieval-Augmented Generation (RAG) system that answers developer questions about the Upwork API. The system reads a PDF, chunks the text, embeds it locally with sentence-transformers, stores vectors in FAISS, and uses the DeepInfra-hosted Meta-Llama-3.1-8B model to generate grounded answers. --- Architecture --- Difficulties Faced • PDF text extraction quality: PyMuPDF (fitz) was chosen over pdfplumber because it better handles multi-column API docs, avoiding scrambled parameter descriptions that would produce misleading chunks. • Chunk boundary problem with code snippets: API documentation...
Skills
- Python