Data & BIpattern
Dual-mode retrieval (full-text + Gemini long-context)
RAG embedding pipelines drift and miss context. Brute-force everything into Gemini 2.5 Flash's 1M context at 3¢/M tokens often wins on quality.
What
Two retrieval algorithms A/B-tested against each other: (a) full-text + Leiden clustering, (b) brute-force dump everything into Gemini 2.5 Flash's million-token context. Skeptical of embeddings/RAG; the long-context option is cheap enough at 3¢/M tokens to win in practice.
← All casesC-024