RAG Retrieval Pipeline Design

Designed the end-to-end RAG pipeline — chunking, embeddings, indexing and grounding — so answers stay accurate, cited and cost-controlled.

Context: LLMs hallucinate without grounding. Decision: a deliberate pipeline — 800-char / 120-overlap chunking, embeddings, vector indexing, and answer synthesis constrained to retrieved context with sources kept. Trade-off: ingestion & tuning effort for grounded, trustworthy answers at controlled token cost.