5.2.4 · Lesson 13 of 14
Retrieval System & Embedding Troubleshooting
Fix RAG failures from bad chunking, missing exact terms, and stale Knowledge Base indexes.
WHAT
Diagnose RAG retrieval failures, poor vector matches, and stale data.
COMMON ISSUES
- Bad chunk boundaries: re-ingest with hierarchical parent-child or layout-aware parsing.
- Exact terms ignored: Hybrid Search (BM25 + k-NN) with reranking.
- New S3 docs missing: run
StartIngestionJobor EventBridge-triggered sync.
Key points
- Hierarchical / layout-aware chunking
- Hybrid Search BM25 + vector
- Reranking
- StartIngestionJob sync
- Stale S3 uploads
Exam focus
Pure vector search misses exact IDs/codes — use Hybrid Search. New S3 files need StartIngestionJob; KB does not auto-reindex by default.