RAG in practice: answer questions from your own documents

Retrieval augmented generation, usually called RAG, is the most common way to make a language model answer questions about your own documents: handbooks, product docs, contracts, support articles. Instead of hoping the model knows the answer, you search your documents first and give the model the relevant passages to answer from.

A RAG demo takes an afternoon. A RAG system that people trust takes much more care. Most failures come from retrieval, not generation: the right passage was never found, was split in half or was buried under near duplicates. This course focuses on getting retrieval right and proving it with measurements.

What you will build

Askdocs, a question answering service over a company knowledge base of a few thousand pages, with citations, permission aware search, fresh content and an evaluation suite.

What you will learn

• When RAG is the right approach, and when search, fine tuning or long context is better
• Parsing and cleaning documents, including PDFs and HTML
• Chunking strategies and keeping structure such as headings
• Embeddings and vector search with PostgreSQL and pgvector
• Hybrid search combining keywords and vectors, and reranking
• Prompting for grounded answers with citations and refusals
• Evaluating retrieval and answers separately
• Access control, freshness, cost and monitoring in production

You should be comfortable with Python, SQL basics and calling an LLM API. Experience with any web framework helps for the final section.