RAG From Zero

Overview

Google, IBM & Meta Certificates — All 10,000+ Courses at 40% Off

One annual plan covers every course and certificate on Coursera. 40% off for a limited time.

RAG from Zero is a hands-on two-module course that teaches you to build production Retrieval-Augmented Generation pipelines in Rust by walking two real tools you can use the same day. Module 1 walks the encode-chunk-index-fuse-retrieve pipeline one stage at a time using the published aprender-rag crate — RecursiveChunker(512, 50) with overlap, MockEmbedder(384) for deterministic teaching with candle for production, reciprocal-rank fusion at k=60, and a closing aprender_film_search demo against a 50-row Sakila fixture that asserts four runtime contracts. Module 2 walks pmat query, a production code-search RAG that ranks by semantic intent plus pagerank plus structural signals — --churn (90-day git volatility), --duplicates (MinHash + Locality-Sensitive Hashing clones), --entropy (pattern diversity), --faults, and -G git-history fusion. The course closes with cross-project search across a sibling-repo workspace via --include-project and --include-source so you can navigate a multi-crate codebase as one indexed corpus. No toy fixtures, no aspirational APIs — aprender-rag is on crates.io today, pmat ships from paiml/pmat, and the companion paiml/rag-from-zero repo runs end-to-end with cargo run and zero infrastructure.

Syllabus

Module 1: aprender-rag — In-Process Text RAG

Build a complete five-stage RAG pipeline (encode → chunk → index → fuse → retrieve) in pure Rust with aprender-rag. You'll wire RecursiveChunker(512, 50) for 50-character overlap that repairs query seams, MockEmbedder(384) for deterministic teaching-grade embeddings (no GPU, no model download, no network), and FusionStrategy::Rrf { k: 60 } for reciprocal rank fusion that lifts long-tail recall without learned weights. The closing demo runs aprender_film_search against a 50-row Sakila film fixture and emits top-5 JSON with four runtime assert! contracts that fire if anything drifts.

Module 2: pmat query — Production Code-Search RAG

Apply the same five-stage RAG pipeline to source code instead of text. The pmat query tool indexes a workspace where chunks are functions, then layers production-grade enrichment on top: search modes (--literal for exact ripgrep-style match, --regex for pattern, semantic by default), enrichment flags (--churn for 90-day Git volatility, --duplicates for MinHash+LSH clone detection, --entropy for diversity, --faults for Batuta unwrap/panic/unsafe annotations, -G for git-history RRF fusion), and the --coverage-gaps mode that ranks every function by uncovered line count so you write tests for the highest-leverage gaps first.