AI 용어집
인공지능 완전 사전
Faithfulness Score
Metric evaluating the degree of faithfulness of the generated response to the provided context, measuring whether the statements are factually supported by the retrieved sources.
Context Relevance
Indicator quantifying the relevance of retrieved documents or passages to the initial query, essential for evaluating the quality of the RAG retrieval component.
Answer Relevance
Score measuring the extent to which the generated response directly and completely answers the posed question, regardless of factual accuracy.
Retrieval Precision
Proportion of relevant documents among all retrieved documents, evaluating the system's effectiveness in returning only useful information.
Retrieval Recall
Ratio of relevant documents retrieved compared to the total number of relevant documents available in the knowledge base.
Knowledge F1 Score
Harmonic mean between precision and recall of retrieved knowledge, providing a balanced measure of the overall RAG system performance.
Context Utilization Rate
Percentage of relevant information from the retrieved context that is actually used in the final response, measuring the efficiency of source utilization.
Hallucination Rate
Frequency at which the model generates information not supported by the provided context, a critical indicator of RAG system reliability.
Semantic Similarity Score
Semantic similarity measure between the generated response and a reference response, using embeddings to capture meaning nuances.
Answer Completeness
Evaluation of the coverage of all relevant aspects of the question in the generated response, ensuring a comprehensive answer.
Retrieval Latency
Time required to retrieve relevant documents from the knowledge base, a crucial criterion for user experience in production.
Token Efficiency Ratio
Ratio between the number of relevant tokens used and the total number of tokens generated, measuring the economic efficiency of the RAG system.
Groundedness Score
Metric assessing the extent to which each statement in the response is supported by explicit evidence in the retrieved sources.
Source Attribution Accuracy
Accuracy with which the system correctly attributes each part of the response to its appropriate documentary source in the retrieved context.
Response Consistency
Measure of the internal consistency of the generated response, assessing the absence of contradictions between different parts of the response.
Query Ambiguity Resolution
Ability of the RAG system to interpret and resolve ambiguities in the user query to retrieve the most relevant information.
Information Overlap Score
Measure of the overlap between information present in the response and that available in the retrieved context, avoiding redundancies.
Answer Accuracy
Evaluation of the factual truthfulness of the generated response compared to a ground truth or validated reference sources.
Retrieval Coverage
Extent of the knowledge base actually accessible by the retrieval system, impacting the ability to answer diverse questions.
Response Coherence
Quality of the logical structure and narrative flow of the generated response, ensuring clear and understandable presentation of information.