Token Count Drives LLM API Costs in Node.js Knowledge Retrieval

The article argues that controlling token usage in a private knowledge-base assistant matters less for prompt brevity than for preserving a stable JSON response contract. It explains that providers bill for both input and output tokens, so irrelevant context and unbounded responses each raise costs. The author recommends embedding-based retrieval, reranking a small candidate set, and keeping whole evidence chunks intact rather than truncating strings. A small answer type with fields for the answer, sources, and a review flag is proposed to stay provider-independent. The piece also suggests a measured experiment varying the number of retrieved chunks while tracking token usage and citation accuracy.