4.4. Reflection Checkpoint

Key Takeaways

• RAG is a two-half pipeline — indexing (chunk → embed → store) offline, query (embed → retrieve → ground) at request time; diagnose failures by stage, not symptom.
• Chunk size is a tuned tradeoff; the query and the index must use the same embedding model; choose Text Split (fixed) vs. Document Layout (structure-aware) by document type.
• Retrieval combines strategies: keyword catches exact terms, vector catches meaning, hybrid (RRF) maximizes recall, semantic ranker (cross-encoder) adds top-result precision — pure vector misses identifiers.
• Indexers + skillsets + integrated vectorization replace hand-written embedding ETL and keep indexes fresh; prefer the Foundry-native grounding path over deprecated "On Your Data."
• Extraction is upstream of retrieval: layout (any doc), prebuilt (standard types), custom (your unique forms) — don't custom-train what a prebuilt covers.
• Grounded agents need citations and document-level security filters; grounding must respect access control.

Connecting Forward

Phase 5 covers computer vision, where a major shift from the predecessor exam appears: multimodal generative models now handle most visual understanding via prompts, reserving custom training for narrow cases. The grounding and modality filters carry straight over — a "describe this image and cite the manual" requirement blends vision understanding with the retrieval you just learned.

Self-Check Questions

• Walk the RAG pipeline for "answer support questions from our product manuals with citations." Name what each stage does and one failure mode per stage.
• A requirement says: retrieve across thousands of policy PDFs, match both exact plan codes and paraphrased concepts, keep results fresh as policies change, and ensure each user only sees their region's documents. Name the search mode, the ingestion mechanism, and the security control — and why no single one of them satisfies the whole requirement.