Companion videos

Chunking, reranking, and hybrid search — companion videos

This article steps past the "throw the docs in a vector DB" version of RAG into the three knobs that actually move quality: chunking, reranking, hybrid search. These videos let you see those knobs being turned. The primary pick is the patient deep-dive into what your splitter is actually doing; the agentic RAG build shows all three knobs turned inside one runnable pipeline.

Primary pick

1:08:59

The 5 Levels Of Text Splitting For Retrieval

Greg Kamradt

The article spends a lot of words on chunking; this is the longest, most patient explanation of what each chunking strategy is actually doing — from character-recursive through document-aware to semantic and agentic splitting. Pair it with Greg's free ChunkViz tool to build intuition before you start tuning.

What you should get from this: Build intuition for chunking choices before tuning a real retrieval system.

Watch or know first: Basic familiarity with embeddings and why retrieval quality depends on how documents are split.

AI Expert note: The specific libraries and demos may age, but the chunking tradeoffs remain stable. Use this for diagnosis language, then test chunk sizes and splitters against your own documents and queries.

Open video page

Also worth watching

24:01

"I want Llama3 to perform 10x with my private knowledge" - Local Agentic RAG w/ llama3

AI Jason

Covers the exact stack the article argues for — query translation, hybrid retrieval, reranking, and a corrective-RAG loop — in one runnable build. Useful as a working mental model for what the chunk → rerank → answer pipeline looks like when it's actually doing its job.

What you should get from this: See how query rewriting, hybrid retrieval, reranking and corrective loops fit into one RAG pipeline.

Watch or know first: Know the basic retrieve-then-generate pattern and be comfortable reading a code walkthrough.

AI Expert note: Treat the framework, model and Llama3-specific setup as version-sensitive. Keep the pipeline shape, but verify current package APIs, model choices, reranker quality and eval results before copying the implementation.

Open video page