Constructing a RAG (Risk Assessment Grid) with Qwen3: A Step-by-Step Guide

In the realm of artificial intelligence (AI), a new development has taken place that might not seem directly related to weather, but is indeed significant. A RAG (Retrieval Augmented Generation) pipeline has been constructed using the Qwen models on a blog website containing approximately 40 blogs.

The script for building the RAG includes several key functions. It reads blog content, chunks it for easier processing, uses FAISS to create a vector store, and employs a reranker to reorder the retrieved documents based on their relevance to the query. The instruct model is also incorporated to provide the final output from the RAG.

The Qwen family, the models at the heart of this project, boast multilingual support and a context length of 256K. They have been developed by Alibaba, the manufacturer of the Qwen3 models, which were released in recent months. Some derivative models, like DeepSeek-R1-0528-Qwen3-8B, are based on Alibaba’s Qwen3 architecture.

The similarity of the same texts is always 1, but the next immediate highest similarity is 0.68, which is between a sentence about AI and a sentence about machine learning (ML). This suggests that while the connection between weather and AI might not be high, the link between AI and ML is more pronounced.

The Qwen3-Reranker model scores each chunk against the query to order the list of documents and assign priority. This helps in delivering the most relevant documents first. The outputs of the RAG seem promising, and for next steps, you could try increasing the documents passed to the instruct model or use a thinking model for a different use case.

Mounish V, a Data Science Trainee with a passion for technology and innovation, led this project. With a deep interest in Deep Learning and Generative AI, Mounish is focused on Data Science and is excited about the possibilities that this new development opens up.

For more information about the documents retrieved, their similarity scores with the query, and the reranker scores, you can refer to the log file 'rag_retrieval_log.txt'. This provides a detailed account of the pipeline's performance.