Recall Conversion Rate (RCR) in RAG

• Recall Conversion Rate (RCR) is defined as the ratio of the F1 score to Recall@k, diagnosing the efficiency of using retrieved evidence for accurate answers.
• RCR reveals that high retrieval recall can include noise and irrelevant content, leading to a 'high recall, low conversion' phenomenon in RAG evaluations.
• NeocorRAG leverages RCR to optimize retrieval quality by filtering noise and enhancing explicit evidence chains, thereby boosting downstream reasoning performance.

Recall Conversion Rate (RCR) is an evaluation metric for Retrieval-Augmented Generation (RAG) that quantifies the contribution of retrieval to reasoning accuracy by taking the ratio of answer accuracy to retrieval coverage, instantiated in "NeocorRAG: Less Irrelevant Information, More Explicit Evidence, and More Effective Recall via Evidence Chains" as dataset-level $\mathrm{F1}/\mathrm{Recall@5}$ and typically reported as a percentage (Peng et al., 30 Apr 2026). It was introduced to diagnose a discrepancy in which improvements in retrieval performance do not consistently translate to commensurate gains in downstream reasoning. In this formulation, RCR serves as a compact diagnosis of whether additional recall is actually useful for reasoning: it does not replace Recall@k, Hit@k, F1, or EM, but conditions end-task accuracy on retrieval coverage and thereby exposes a "high recall, low conversion" phenomenon that recall-centric evaluation can miss.

1. Motivation and diagnostic role

RAG has long leaned on recall-centric retrieval metrics such as Recall@k, NDCG, and MRR to signal progress. The motivation for RCR is the observation that retrieval coverage alone does not ensure that the retrieved set is usable for reasoning. A retrieved package may contain the needed support while also containing misleading snippets or merely irrelevant ones, and these cases can produce divergent QA outcomes despite identical recall. Traditional retrieval metrics indicate whether relevant items were surfaced; they do not indicate whether the context is non-misleading, sufficiently explicit, or supportive of reasoning chains (Peng et al., 30 Apr 2026).

The paper frames the core empirical symptom as a “high recall, low conversion” phenomenon. A concrete example given in the study is that HippoRAG2 reaches passage Recall@5 of 96.3% on HotpotQA while still attaining only approximately 75.5% F1. The central claim is that high retrieval coverage can coexist with substantial unrecovered downstream accuracy because recall gains are often achieved by admitting more irrelevant or interfering content. This suggests that raw recall is an incomplete optimization target when the end objective is reasoning rather than retrieval in isolation.

2. Formal definition

The paper defines RCR in Section 3.1 as follows:

$$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$

Here, $\mathcal{Q}$ is the set of queries, $A$ and $A^*$ denote the generated and ground-truth answers, and $\mathrm{F1}(A,A^*)$ is the dataset-level answer F1 averaged over queries. For each query $q$, $\mathcal{D}_g(q)$ is the set of supporting passages. The retrieval term is instantiated as dataset-level $\mathrm{Recall@}n$, where the per-query $\mathrm{Recall@}n(q)$ is the fraction of $$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$0 appearing in the top-$$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$1 retrieved passages, and the dataset-level $$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$2 is the average of these per-query values.

Unless otherwise stated, the study uses $$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$3 and reports $$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$4. In the tables, RCR is shown as a percentage, that is, $$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$5. The paper gives BM25 on NQ as an example: $$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$6. If dataset-level recall is zero, then RCR is defined to be zero. No additional variants—instance-level, chain-level, or multi-hop-specific—are formally defined; all reported results use the dataset-level ratio with Recall@5 (Peng et al., 30 Apr 2026).

3. Computation and evaluation protocol

Computing RCR requires only standard RAG artifacts. For each query, the necessary inputs are the ground-truth answer, an answer scorer for $$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$7, the ground-truth supporting passages $$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$8, the top-$$\mathrm{RCR}(\mathcal{Q}, \mathcal{D}_g, A, A^*)= \begin{cases} \dfrac{\mathrm{F1}(A, A^*)}{\mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})}, & \text{if } \mathrm{Recall}(\mathcal{D}_g,\mathcal{Q})>0 \ 0, & \text{otherwise.} \end{cases}$$9 retrieved items $\mathcal{Q}$0, and the predicted answer $\mathcal{Q}$1. Evidence chains or chain annotations are not used to compute RCR; they are used by NeocorRAG to optimize retrieval quality.

The metric handles single-hop and multi-hop settings uniformly. If a query has multiple relevant documents, per-query recall is the fraction of supporting passages surfaced in $\mathcal{Q}$2. If $\mathcal{Q}$3 and $\mathcal{Q}$4 supports are retrieved, then $\mathcal{Q}$5. Partial evidence in multi-hop QA therefore yields partial recall directly. Chain correctness is not part of the definition: RCR measures conversion of retrieval coverage into answer accuracy irrespective of whether the model explicitly follows the correct chain.

Aggregation is performed at the dataset level. One computes the mean F1 over queries, the mean Recall@k over queries, and then forms $\mathcal{Q}$6 as either a fraction or, as reported in the paper, as $\mathcal{Q}$7. The recommended setting in the study is $\mathcal{Q}$8. If some queries have no annotated supports, the paper notes that one may exclude them or define $\mathcal{Q}$9 and let F1 contribute as usual; the experiments use benchmarks with annotated supports (Peng et al., 30 Apr 2026).

The worked example in the paper makes the aggregation explicit. Suppose a dataset has three questions at $A$0: for $A$1, $A$2 and both are retrieved, so $A$3 and $A$4; for $A$5, $A$6 and one is retrieved, so $A$7 and $A$8; for $A$9, $A^*$0 and it is retrieved, so $A^*$1 and $A^*$2. The dataset Recall@5 is approximately $A^*$3, the dataset F1 is $A^*$4, the fractional RCR@5 is approximately $A^*$5, and the percentage form is $A^*$6.

4. Empirical behavior on question-answering benchmarks

The principal empirical finding is that, across mainstream RAG methods, RCR exhibits a near-linear decay as Recall@5 increases. The paper presents this as a qualitative, figure-based result and does not report regression coefficients or correlation values. The interpretation offered is that many recall improvements are achieved by admitting more noise and interfering content, which reduces the effectiveness of the retrieved context for reasoning (Peng et al., 30 Apr 2026).

On the four benchmarks NQ, MuSiQue, 2Wiki, and HotpotQA, results are reported for Llama-3.3-70B and Llama-3.2-3B answerers. With the 70B answerer, HippoRAG2 attains average $A^*$7, $A^*$8, and $A^*$9. NeocorRAG, using identical retrieval inputs from HippoRAG2, attains average $\mathrm{F1}(A,A^*)$0, $\mathrm{F1}(A,A^*)$1, and $\mathrm{F1}(A,A^*)$2. On HotpotQA specifically, HippoRAG2 records $\mathrm{F1}(A,A^*)$3, $\mathrm{F1}(A,A^*)$4, and $\mathrm{F1}(A,A^*)$5, whereas NeocorRAG records $\mathrm{F1}(A,A^*)$6, $\mathrm{F1}(A,A^*)$7, and $\mathrm{F1}(A,A^*)$8. On NQ, the comparison is $\mathrm{F1}(A,A^*)$9 for HippoRAG2 versus $q$0 for NeocorRAG; on MuSiQue, $q$1 versus $q$2; on 2Wiki, $q$3 versus $q$4.

With the 3B answerer, HippoRAG2 attains average $q$5, $q$6, and $q$7, while NeocorRAG attains average $q$8, $q$9, and $\mathcal{D}_g(q)$0. On HotpotQA in the 3B setting, F1 rises from $\mathcal{D}_g(q)$1 to $\mathcal{D}_g(q)$2 and RCR rises from $\mathcal{D}_g(q)$3 to $\mathcal{D}_g(q)$4. The paper summarizes the same-initial-retrieval comparison as average gains of $\mathcal{D}_g(q)$5 RCR points and $\mathcal{D}_g(q)$6 F1 points for 70B, and $\mathcal{D}_g(q)$7 RCR points and $\mathcal{D}_g(q)$8 F1 points for 3B, across the four datasets. It further notes that on HotpotQA the gains reach up to $\mathcal{D}_g(q)$9 RCR points and up to $\mathrm{Recall@}n$0 F1 points under identical prompts.

A complementary filtering experiment reinforces the interpretation of RCR as a retrieval-quality metric rather than a raw-coverage metric. After evidence-chain-based filtering in the 3B setting, average Recall@5 drops slightly from $\mathrm{Recall@}n$1 to $\mathrm{Recall@}n$2, but average RCR@5 rises from $\mathrm{Recall@}n$3 to $\mathrm{Recall@}n$4, while Average Irrelevant Documents decline from $\mathrm{Recall@}n$5 to $\mathrm{Recall@}n$6. This is presented as direct evidence that reducing redundancy and interference can improve conversion efficiency even when raw recall decreases.

5. Retrieval quality optimization and NeocorRAG

Within the paper, RCR is not only an evaluation metric but also the central diagnostic that motivates NeocorRAG. The framework is described as a holistic retrieval quality optimization method that systematically mines and utilizes Evidence Chains. Its pipeline has three components: activated search for a refined candidate evidence space, constrained decoding for precise evidence chain generation, and evidence-guided document filtering. The activated search stage scores candidate paths using overall semantic relevance, a length penalty, and a reward for high-confidence triples. The constrained decoding stage organizes candidate chains in a prefix tree and restricts decoding to prefixes that occur in that tree, thereby preventing hallucinated chains. The filtering stage retains low-confidence documents only when they contain a triple from a generated chain, so that the final context preserves strong items while salvaging only chain-supported supplemental documents (Peng et al., 30 Apr 2026).

The paper positions this design against two failure modes. Structure-enhanced methods such as HippoRAG2 excel at Recall@5 but tend to underperform in RCR because noise and interfering content hinder reasoning. Reasoning-enhanced methods such as Trace and CoRAG suppress noise but often harm recall or rely on model-dependent policies, which yields unstable or lower overall performance. In Table 1, Trace’s average RCR@5 is below NeocorRAG in both 3B and 70B settings. The claimed contribution of NeocorRAG is to raise RCR without sacrificing recall by surfacing explicit, faithful evidence and simultaneously reducing noise.

The paper also emphasizes efficiency. NeocorRAG is described as a training-free paradigm for RAG enhancement and uses under 20% of the tokens of comparable methods overall. Figure 1 is summarized more specifically as showing average token use equal to 20.1% of CoRAG’s and under 2.94% of Trace’s. A plausible implication is that RCR can function as a metric for evaluating not only retrieval effectiveness but also the efficiency of retrieval-quality optimization when comparable answer quality is sought under token constraints.

6. Relation to other metrics, limitations, and open questions

RCR is designed to be complementary to established retrieval and QA metrics. Recall@k and Hit@k measure whether relevant items are present; they do not measure whether the retrieved context is usable or non-misleading for reasoning. Precision, MRR, and NDCG reflect ranking quality or concentration of relevance but still do not capture reasoning impact, especially in multi-document reasoning where subtle interference can dominate. F1 and EM assess answer quality but conflate retrieval effects with parametric knowledge, since a model may answer correctly from internal knowledge rather than from retrieved context. RCR addresses this by normalizing answer accuracy by retrieval coverage and is said to be most revealing in high-recall regimes and multi-hop tasks, where over-retrieval can yield noise or misleading snippets (Peng et al., 30 Apr 2026).

The metric also has explicit limitations. Its denominator depends on gold support annotations, so incomplete or noisy supports can bias recall and therefore bias RCR. When recall is very small, the ratio can be unstable; the paper handles the zero-recall case by defining RCR to be zero. RCR is intentionally model-agnostic and dataset-level, which makes it broadly applicable, but it does not reveal whether a system used the “right” chain or which hop failed in a multi-hop problem. The paper notes that chain-conditioned variants could be explored in future work, but such variants would require reliable chain annotations.

For evaluation practice, the study recommends reporting RCR@5 alongside Recall@5 and F1, especially when comparing against recall-heavy baselines. It also advises monitoring RCR while tuning $\mathrm{Recall@}n$7, because larger $\mathrm{Recall@}n$8 can inflate recall while lowering RCR due to noise. Where support annotations are incomplete, reporting both RCR and raw F1 and supplementing them with qualitative audits is suggested. The statistical reporting in the paper averages over five runs and marks significance at $\mathrm{Recall@}n$9 against Trace in F1, but does not provide confidence intervals for RCR. This suggests that, although RCR is a compact and informative diagnostic, it remains most informative when used together with raw retrieval, answer-quality, and qualitative evidence analyses.