Recall
A dimension for evaluating rationale quality requiring that all necessary rationales are covered without omitting key conditions.
Timeline (last 90 days)
In the RAG pipeline, the recall stage aims for "missing nothing," typically using vector similarity to quickly screen out dozens to hundreds of candidate passages; the reranking stage uses a heavier model for fine-grained scoring, aiming for "accurate ordering."
Long-context capability can be decomposed into three categories: recall, state tracking and pattern learning.
Existing long-context benchmarks focus almost entirely on recall capability, neglecting state tracking and pattern learning.
On LEAPS, models with near-perfect performance on recall tasks may diverge significantly on state tracking and pattern learning, with degradation becoming more severe as input length increases.
Hermes's memory system can be broken down into three reusable concepts: storage, injection, and recall.
A reranker can only reorder within the candidate set already recalled; it cannot pull back a correct answer that never entered the top K.
The goal of the retrieval stage is to quickly filter thousands of relevant candidates from a massive pool (e.g., billions), prioritizing high recall and low latency.
All Facts (7)
In the RAG pipeline, the recall stage aims for "missing nothing," typically using vector similarity to quickly screen out dozens to hundreds of candidate passages; the reranking stage uses a heavier model for fine-grained scoring, aiming for "accurate ordering."
50%UnverifiedLong-context capability can be decomposed into three categories: recall, state tracking and pattern learning.
50%UnverifiedExisting long-context benchmarks focus almost entirely on recall capability, neglecting state tracking and pattern learning.
50%UnverifiedOn LEAPS, models with near-perfect performance on recall tasks may diverge significantly on state tracking and pattern learning, with degradation becoming more severe as input length increases.
50%UnverifiedHermes's memory system can be broken down into three reusable concepts: storage, injection, and recall.
50%UnverifiedA reranker can only reorder within the candidate set already recalled; it cannot pull back a correct answer that never entered the top K.
50%UnverifiedThe goal of the retrieval stage is to quickly filter thousands of relevant candidates from a massive pool (e.g., billions), prioritizing high recall and low latency.
50%