Vector Institute’s RAG Bootcamp: Six Reference Implementations You Can Actually Run

Vector Institute published the reference code for their RAG Bootcamp (Nov 2024–Jan 2025) as a single repository with six self-contained implementations. Each targets a distinct retrieval pattern: web search for live knowledge, document search over unstructured PDFs, text-to-SQL over a Portuguese banking dataset, cloud object storage (AWS S3) retrieval, a full PubMed biomedical pipeline with reranking, and a dedicated evaluation module built on the Ragas framework. The stack is intentionally mainstream — Python 3.10+, LangChain, LlamaIndex — so the patterns transfer directly to production codebases already using those libraries. What the implementations cover Web Search: Demonstrates plugging a search API into an LLM to answer questions beyond training cutoff (e.g., 2024 World Series winner). Document Search: End-to-end ingestion, chunking, embedding, vector storage, and retrieval over domain documents. SQL Search: Natural language to SQL over structured relational data, using a real Kaggle banking dataset. Cloud Search: Retrieval from AWS S3 objects, showing how to bridge object storage into a RAG pipeline. PubMed QA: The most complete pipeline — ingestion through evaluation — on biomedical literature with explicit reranking and metric tracking. RAG Evaluation: Standalone notebooks applying Ragas test sets to measure faithfulness, answer relevance, context