NeocorRAG: Enhanced Retrieval-Augmented Generation

Updated 4 July 2026

NeocorRAG is a retrieval-augmented generation framework that restructures evidence chains to convert high recall into improved downstream reasoning.
It introduces the Recall Conversion Rate (RCR) metric to measure how efficiently retrieved support translates to higher answer quality.
The training-free, plug-and-play design leverages activated search, constrained decoding, and evidence-guided context refinement to outperform traditional RAG methods.

NeocorRAG is a retrieval-augmented generation framework that addresses a specific gap in RAG system design: improvements in retrieval performance do not consistently translate into commensurate gains in downstream reasoning. It formalizes this gap through the Recall Conversion Rate (RCR), a metric intended to quantify how efficiently retrieval coverage is converted into answer quality, and it organizes retrieval around explicit evidence chains rather than only larger or more comprehensive passage sets. The framework is described as training-free, plug-and-play, and focused on holistic retrieval quality optimization through activated search, constrained decoding, and evidence-guided context refinement (Peng et al., 30 Apr 2026).

1. Conceptual basis and motivating problem

NeocorRAG begins from the observation that standard RAG evaluation often privileges retrieval quantity, typically through metrics such as Recall@k, while under-specifying retrieval quality as experienced by the downstream reasoner. In the reported diagnosis, strong graph-based systems can retrieve supporting passages at very high rates without achieving commensurate QA performance. The paper gives a representative case on HotpotQA in which HippoRAG2 attains 96.3% passage Recall@5 but only 75.5% QA F1, indicating that the supporting material is frequently present while reasoning still fails (Peng et al., 30 Apr 2026).

This discrepancy is used to distinguish two dimensions. Retrieval quantity concerns whether gold-support passages appear in the retrieved set. Retrieval quality concerns whether the retrieved content is actually usable for correct reasoning, including its noise level, sufficiency, and the degree to which supporting relations are made explicit. The paper argues that mainstream methods fall into a trade-off pattern. Structure-enhanced RAG methods, including GraphRAG, HippoRAG, and RAPTOR, emphasize comprehensiveness and high recall but often deliver large amounts of extra text without explicit evidence traces. Reasoning-enhanced methods, including TRACE, RetroLLM, IRCoT, and CoRAG, emphasize noise suppression and evidence extraction but may sacrifice recall or require expensive training (Peng et al., 30 Apr 2026).

NeocorRAG is positioned as an attempt to resolve that trade-off by maintaining high recall while optimizing retrieval quality through evidence chain mining. This suggests that the framework treats retrieval not as a terminal ranking stage but as an intermediate structure-construction problem in which explicit, faithful, multi-hop support paths are the central retrieval unit.

2. Recall Conversion Rate and retrieval quality criteria

The central diagnostic metric is Recall Conversion Rate (RCR), defined over a query set $\mathcal{Q}$ , gold supporting document sets $\mathcal{D}_g$ , generated answers $A$ , and gold answers $A^*$ as:

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

In the reported experiments, this is instantiated as RCR@5, pairing token-level answer F1 with Recall@5 over gold supporting passages (Peng et al., 30 Apr 2026).

RCR is intended to separate two failure modes that are conflated by raw F1 or raw Recall@k alone. Low recall indicates retriever failure. Low RCR at high recall indicates conversion failure: the correct material is present, but the model either ignores it, is distracted by noise, or cannot exploit latent multi-hop relations. The paper reports that across mainstream methods, as Recall@5 increases, RCR decreases roughly linearly, described as a “high recall, low conversion” pattern (Peng et al., 30 Apr 2026).

To operationalize retrieval quality, the framework classifies retrieved content into four categories:

Irrelevant information: unrelated content.
Interfering information: plausible but incorrect or misleading content.
Explicit evidence: directly supportive content.
Hidden evidence: relevant content whose utility depends on associative or multi-step reasoning.

From this taxonomy, NeocorRAG defines three optimization criteria: More Comprehensive, Less Noise Information, and Evidence Visualization. The first preserves useful evidence, including hidden associations; the second suppresses irrelevant and interfering material; the third makes evidence chains explicit and interpretable for the QA model (Peng et al., 30 Apr 2026). A plausible implication is that NeocorRAG treats retrieval quality as a joint function of coverage, denoising, and representational explicitness rather than as a pure ranking objective.

3. End-to-end architecture and evidence-chain pipeline

NeocorRAG is described as a training-free, plug-and-play enhancement on top of an existing retriever, instantiated in the paper with HippoRAG2. Its end-to-end pipeline has three stages: Candidate Evidence Space Identification (Activated Search), Precise Evidence Chain Generation (Constrained Decoding), and Retrieved Context Optimization (Evidence-Guided Filtering) (Peng et al., 30 Apr 2026).

In the first stage, retrieved documents are represented as a graph. The framework identifies query-relevant nodes and runs a dynamic, activation-based path search to obtain candidate evidence chains. In the second stage, those candidate paths are organized into a prefix tree, and a lightweight LLM decodes evidence chains subject to prefix-tree constraints, so that only sequences corresponding to valid paths in the candidate set can be produced. In the third stage, the selected evidence chains are used to filter and augment the document set, yielding a refined context that is both smaller and more tightly aligned with the mined support structure (Peng et al., 30 Apr 2026).

The final QA model receives two kinds of input: the optimized document set and the explicit evidence chains, presented in prompts as “information association paths.” The paper states that the Q&A prompt encourages the model to use these chains in reasoning. This architecture differs from a standard passage-ranking pipeline because the principal output of intermediate retrieval is not only a list of passages but a structured set of cross-document support paths.

The framework’s emphasis on explicit evidence has an important parallel in multilingual RAG work on cultural alignment. CORAL, for example, uses an adaptive retrieval loop that critiques evidence for relevance and cultural alignment, then revises both retrieval space and query formulation when evidence is insufficient (Lee et al., 28 Apr 2026). While the two systems target different failure modes, both treat retrieval as an iterative control problem rather than a fixed top- $k$ selection problem. This suggests a broader trend in RAG research toward explicit retrieval-condition optimization rather than passive context accumulation.

4. Activated search and constrained decoding

The activated search algorithm is the core of NeocorRAG’s first stage. Starting from a graph $G=(V,E)$ built over the retrieved documents, the method first extracts key entities from the query using an LLM and identifies initial nodes by embedding similarity with threshold $\tau_{\text{node}}$ , reported as 0.9 in experiments (Peng et al., 30 Apr 2026).

From these activated nodes, the system performs constrained Depth-First Search with a maximum hop limit $K$ . At each step, it forms a triple $t=(v_i,r,n_j)$ between a current node, relation, and neighbor, computes a local similarity score between the triple embedding and the query embedding, and only expands neighbors whose score exceeds an edge threshold $\mathcal{D}_g$ 0 (Peng et al., 30 Apr 2026). This early pruning is intended to prevent semantically weak edges from propagating into long, noisy paths.

Candidate paths are then assigned a global score combining three factors: a length penalty, overall semantic relevance, and a high-confidence activation reward. The expected maximum path length is reported as 10 in experiments, and the high-confidence reward is applied when a triple-query similarity exceeds $\mathcal{D}_g$ 1, where $\mathcal{D}_g$ 2 (Peng et al., 30 Apr 2026). The design rationale is explicit: long paths beyond the expected limit are penalized, semantically aligned paths are rewarded, and paths with multiple highly relevant triples receive multiplicative reinforcement.

The second stage converts candidate paths into final evidence chains through prefix-tree-constrained decoding. Candidate chains are linearized and inserted into a trie. A lightweight LLM, such as Llama 3.2 3B or Qwen2.5, then decodes evidence sequences under the constraint that every emitted prefix must correspond to a valid trie prefix (Peng et al., 30 Apr 2026). The constraint function ensures that the model cannot hallucinate unsupported edges or tokens. Every generated chain therefore corresponds to a real path in the graph induced by the retrieved documents.

This stage uses beam search to generate multiple diverse evidence chains, reflecting the fact that multi-hop questions often require several complementary chains (Peng et al., 30 Apr 2026). The paper frames the LLM’s role here not as inventing new evidence but as re-ranking and selecting among faithful candidates on the basis of global reasoning.

5. Evidence-guided filtering, context refinement, and training-free integration

The third stage uses evidence chains to optimize the retrieved context. Let the initial document set be $\mathcal{D}_g$ 3, partitioned into $\mathcal{D}_g$ 4, the top- $\mathcal{D}_g$ 5 highest-ranked documents, and $\mathcal{D}_g$ 6, the lower-ranked remainder. Let $\mathcal{D}_g$ 7 denote the union of triples appearing in the selected evidence chains. NeocorRAG defines a supplement set:

$\mathcal{D}_g$ 8

and then forms the refined context as:

$\mathcal{D}_g$ 9

This filtering procedure preserves the highest-ranked documents by default while “rescuing” lower-confidence documents only when they contain triples that appear in selected evidence chains (Peng et al., 30 Apr 2026). The intended effect is twofold: a large fraction of irrelevant low-confidence documents are removed, while hidden but important material is retained if it participates in explicit support structure.

The framework is explicitly characterized as training-free with respect to large components. The base retriever, encoders, and QA LLMs remain fixed; the optimization occurs through prompting, algorithmic search, and constrained decoding rather than gradient-based retraining (Peng et al., 30 Apr 2026). This is presented as a practical advantage for adoption: the method can be inserted into existing RAG systems so long as the retriever can deliver a graph-like structure or triples.

The paper also emphasizes model-agnostic deployment. It reports using bge-large-en-v1.5 as encoder, Llama-3.2-3B-Instruct and Llama-3.3-70B-Instruct as QA LLMs, and lightweight Llama 3 or Qwen2.5 models from 1B to 14B for evidence-chain mining (Peng et al., 30 Apr 2026). This suggests that the framework’s main contribution lies in retrieval-time restructuring rather than in a specific LLM scale.

6. Empirical performance, ablations, and efficiency

NeocorRAG is evaluated on HotpotQA, 2WikiMultiHopQA, MuSiQue, and Natural Questions (NQ). The baselines include BM25, Contriever, GTR (T5-base), structural RAG systems such as RAPTOR, HippoRAG, and HippoRAG2, and reasoning-enhanced systems such as CoRAG and TRACE (Peng et al., 30 Apr 2026).

With Llama-3.3-70B as QA model, the reported average performance against HippoRAG2 is as follows:

Method	Avg Recall@5	Avg F1	Avg RCR@5
HippoRAG2	84.9	64.6	75.8
NeocorRAG (3B evidence miner)	84.5	66.1	77.9
NeocorRAG (8B evidence miner)	84.5	68.1	80.3

On HotpotQA specifically, HippoRAG2 reports Recall@5 = 96.3, F1 = 75.5, and RCR@5 = 78.4, while NeocorRAG (8B evidence miner) reports Recall@5 = 96.4, F1 = 78.3, and RCR@5 = 81.2 (Peng et al., 30 Apr 2026). Under the 3B QA model, the paper states that improvements are larger, reaching up to +9.4 F1 and +9.6 RCR on HotpotQA.

Against TRACE with a 70B QA model, the paper reports avg F1 = 56.9 and RCR = 66.4 for TRACE using an 8B mining model, versus avg F1 = 68.1 and RCR = 80.3 for NeocorRAG using an 8B mining model, yielding +11.2 F1 and +13.9 RCR on average (Peng et al., 30 Apr 2026).

The token-efficiency claim is also explicit. NeocorRAG uses about 20.1% of the tokens used by CoRAG and <3% of the tokens used by TRACE on average across datasets (Peng et al., 30 Apr 2026). This efficiency is attributed to two design choices: restricting the final context to $A$ 0 and expressing important cross-document connections via compact evidence chains rather than larger raw passages.

Ablation studies on HotpotQA with Llama-3.2-3B QA isolate the contribution of each component. Full NeocorRAG reports EM 53.7, F1 66.8, and RCR@5 69.3. Replacing activated search with naive chain search reduces F1 to 49.5 and RCR to 51.3. Removing constrained decoding yields F1 58.7 and RCR 61.0. Removing document filtering yields F1 57.0 and RCR 59.1 (Peng et al., 30 Apr 2026). These results are used to argue that candidate-space quality, faithfulness constraints, and noise removal are all essential rather than interchangeable.

The effect of filtering is quantified further for the 3B QA model using ATD (Average Total Documents) and AID (Average Invalid Documents among those). Before filtering, the paper reports ATD = 5.0, AID ≈ 2.83, Recall@5 = 84.53, and RCR@5 = 53.48. After evidence-chain-based filtering, it reports ATD = 3.52, AID = 1.39, Recall@5 = 80.61, and RCR@5 = 61.81 (Peng et al., 30 Apr 2026). The reported interpretation is that a minor decrease in recall is offset by a large improvement in conversion efficiency, with more than half of irrelevant documents removed.

7. Position in the RAG landscape, evaluation, and limitations

NeocorRAG is situated between two established lines of RAG research. Relative to pure recall-oriented methods, it accepts that high recall is necessary but argues that retrieval quality must be optimized explicitly. Relative to pure quality-oriented methods, it seeks to avoid aggressive pruning that discards useful supporting evidence. Its contribution is therefore framed as a holistic view of retrieval quality centered on maintaining sufficient coverage, eliminating interfering noise, and making evidence explicit and usable (Peng et al., 30 Apr 2026).

The emphasis on explicit evidence traces also bears on evaluation. A separate evaluation framework, CCRS, proposes five zero-shot LLM-as-a-judge metrics—Contextual Coherence (CC), Question Relevance (QR), Information Density (ID), Answer Correctness (AC), and Information Recall (IR)—for end-to-end RAG assessment (Muhamed, 25 Jun 2025). CCRS is not part of NeocorRAG itself, but it is directly relevant to systems whose changes affect both retrieval and reasoning. In particular, CCRS distinguishes answer grounding relative to retrieved context from correctness relative to gold answers, and it is reported as about 5× faster than RAGChecker in the cited experiments (Muhamed, 25 Jun 2025). This suggests that NeocorRAG-like systems may benefit from evaluation protocols that can separately track denoising, grounding, completeness, and correctness rather than relying only on recall and answer F1.

The paper states several limitations. NeocorRAG depends on the base retriever and graph quality: if the underlying retriever fails to surface key evidence, the framework cannot recover it. The constraints may be brittle if the candidate evidence space contains noisy triples or poor edges. There is also a scalability question for very large corpora or graphs, even though activated search operates on subgraphs derived from retrieved documents. Finally, despite its token efficiency, the multi-stage pipeline introduces its own latency through graph search, constrained decoding, and QA generation (Peng et al., 30 Apr 2026).

Future directions mentioned or implied include better automatic evaluation metrics for evidence chains themselves, end-to-end trainable variants that preserve structural faithfulness, and applying evidence-chain mining beyond QA to fact checking, multi-document summarization, and scientific reasoning (Peng et al., 30 Apr 2026). A plausible implication is that NeocorRAG is less a single-task heuristic than a general retrieval-time design pattern: organize evidence into faithful relational paths, then use those paths to determine what context should remain visible to the generator.