Definition
A configurable numerical cutoff applied to the similarity scores of retrieved document chunks in a RAG pipeline to exclude results that do not meet a minimum mathematical alignment with the user's query. It acts as a noise filter to prevent irrelevant context from being injected into the LLM's prompt, thereby reducing hallucinations and token waste.
Mathematical vector distance cutoff vs. general topical importance.
"A high-jump bar in an Olympic event; only the athletes (data chunks) who can clear the specific height are allowed to move on to the final round (the LLM prompt)."
- Cosine Similarity(Prerequisite)
- Top-K Retrieval(Component)
- Hallucination Mitigation(Objective)
- Precision-Recall Trade-off(Architectural Trade-off)
Conceptual Overview
A configurable numerical cutoff applied to the similarity scores of retrieved document chunks in a RAG pipeline to exclude results that do not meet a minimum mathematical alignment with the user's query. It acts as a noise filter to prevent irrelevant context from being injected into the LLM's prompt, thereby reducing hallucinations and token waste.
Disambiguation
Mathematical vector distance cutoff vs. general topical importance.
Visual Analog
A high-jump bar in an Olympic event; only the athletes (data chunks) who can clear the specific height are allowed to move on to the final round (the LLM prompt).