What you should be able to do
Design a production retrieval pipeline, choose chunking and ranking strategies, cite sources, and debug hallucinations caused by poor retrieval.
CapstoneBuild a cited technical-support assistant over a document library.
Ground models in external knowledge
Retrieval-Augmented Generation
RAG combines a generator with a retriever so an AI system can look up relevant information instead of relying only on weights learned during training.
User question
Embed query
Retrieve documents
Rerank evidence
Answer with citations
Concepts to master
- Embeddings and vector search
- Chunking and metadata
- Hybrid retrieval and reranking
- Grounded generation and citation
- Evaluation for answer faithfulness
Builder path
How to turn this topic into a working project.
Use this as the bridge from reading to implementation. The goal is to build a small, inspectable version before adding frameworks or production complexity.
- Create a small docs folder in VS Code and write an ingestion script that loads each file with its URL and title.
- Split the documents into chunks, attach metadata, and store embeddings in a local or managed vector index.
- Build a query route that retrieves evidence, assembles a citation prompt, and refuses when evidence is missing.
- Write retrieval and answer-evaluation cases before changing chunk size, embedding model, reranker, or prompt rules.
Primary sources
Back to roadmap Open first source