Chunking Strategies for RAG
Problem statement
A team is building retrieval-augmented generation (RAG) over two corpora: a library of annual and quarterly financial reports (long PDFs with many tables and footnotes, near-identical text from one period to the next), and an enterprise knowledge base that mixes chat threads, issue-tracker tickets, shared documents and PDFs, each with its own access permissions. Retrieval is judged by recall@5 on a labeled set of questions. Work through the graded checks, then answer each interview question as you would out loud.
Show hint
Ask what unit of text would answer a typical question on its own.
Relevant roles: AI Engineer, Gen AI Engineer and Forward Deployed Engineer.