RAG-Based Few-Shot Structured Document Generation Pipeline
May 2026—Present- Architected a RAG-driven few-shot document generation pipeline, using
nomic-embed-textvector search to dynamically inject relevant template examples into Claude API calls. - Engineered a zero-overhead vector cache system using serialized flat-file storage (
pickle), leveraging NumPy to eliminate redundant embedding calculations. - Reduced operational costs by 4.8% by implementing a Held-Karp TSP optimization layer for prompt sequencing to maximize LLM context caching.