NVIDIANVIDIA-Certified Professional: Agentic AI2026-106.3 Build extract, transform, and load (ETL) pipelines to integrate enterprise or client data sources.
Quickstart: retriever CLI
Completed by opening the official lab
Why it matters
Extraction quality decides what retrieval can find, and NeMo Retriever runs it as a pipeline of NIMs rather than one parser. This quickstart puts each stage behind a flag, so you can see which models touch a document.
What to practise
Run retriever ingest on a PDF and retriever query against the default LanceDB table, then scale out with retriever ingest batch and its worker flags, or send the corpus to a service with retriever ingest service. Route extraction to hosted NIMs with --page-elements-invoke-url, --ocr-invoke-url and --table-structure-invoke-url. When you query, set --candidate-k above --top-k and filter with --content-types table, and watch which hits survive. Then try --agentic, which issues several sub-queries and fuses the candidates.
How it fits
The page-elements, OCR and table-structure stages are 6.3.1’s extraction by modality; compare what the content-type filter returns with 6.3.1’s structure that retrieval needs. The gap between --candidate-k and --top-k is 6.1.3’s candidate pool, and --agentic is a small form of 6.1.4’s query decomposition.