Welcome to the future of AI-assisted knowledge retrieval! This repository showcases a hybrid RAG (Retrieval-Augmented Generation) system powered by LLaMA 3.1, offering an unparalleled combination of local data processing and optional web-based search.
This project implements a sophisticated workflow for answering user queries by:
1️⃣ Searching local documents for relevant information.
2️⃣ Falling back on web search when local data isn’t enough.
3️⃣ Using the LLaMA 3.1 model to generate concise and insightful answers.
The system ensures data privacy while maintaining flexibility by integrating secure, local processing with web-enhanced intelligence.
- 💾 Local Data First: Process user-provided PDFs or URLs to create a custom knowledge base.
- 🌐 Web Search Fallback: Dynamically fetch answers from the web if local data is insufficient.
- 🧠 Powered by LLaMA 3.1: Generate human-like responses grounded in the retrieved context.
- ⚡ Intelligent Retrieval: Combines document splitting, embedding, and grading to ensure high-quality information retrieval.
- 🛡️ Privacy Focused: Local processing guarantees your data stays secure, with web search used only when necessary.
Upload URLs or PDFs, which are processed as follows:
- Loading: WebBaseLoader fetches online articles, or PDFs are converted to text.
- Splitting: Text is split into smaller chunks using LangChain’s
RecursiveCharacterTextSplitter. - Embedding:
Nomic-Embed-Textembeddings transform text chunks into vector representations.
When a query is submitted:
- Local Search: The query is embedded and matched against the vector database for relevant text chunks.
- Grading: A grader filters out irrelevant results, ensuring only the best information is used.
The relevant chunks are passed to LLaMA 3.1, which generates a concise and insightful response.
If no local data is relevant, the system uses a web search tool (TavilySearchResults) to fetch additional context. These results are processed and added to the retrieved data for answer generation.
An end-to-end pipeline combining local retrieval and web search fallback for comprehensive question answering.
- Enterprises: Securely process internal documents while integrating external knowledge when necessary.
- Researchers: Combine custom datasets with real-time web resources for deep insights.
- AI Enthusiasts: Experiment with advanced RAG systems combining local and web-based retrieval.
-
Retrieval-Augmented Generation (RAG):
- Combines two paradigms:
- Retrieval Models: Efficiently search for relevant documents.
- Generative Models: Generate coherent and contextually rich answers.
- Combines two paradigms:
-
LLaMA 3.1:
- A state-of-the-art language model designed for generative tasks.
- Fine-tuned for local inference, reducing reliance on external APIs.
-
Vector Embeddings:
- Maps text into numerical vectors in a high-dimensional space.
- Enables semantic searches based on content similarity.
- Language Model: LLaMA 3.1
- Embedding Generator:
nomic-embed-text-v1.5 - Vector Storage: SKLearn VectorStore
- Text Splitting: LangChain’s RecursiveCharacterTextSplitter
- Web Search: TavilySearchResults
- Programming Language: Python
This project is just the beginning! Planned future enhancements include:
- Multi-language Support: Expand the system to handle non-English documents.
- OCR Integration: Enable document ingestion from scanned images.
- Enhanced Grading: Improve retrieval relevance with advanced scoring mechanisms.
- User Interface: Build a web-based UI for seamless interaction.
Contributions are welcome! If you have ideas or improvements, feel free to open an issue or submit a pull request.
This project is licensed under the MIT License. See the LICENSE file for details.