Criterion-Conditioned RAG Module

• Criterion-Conditioned RAG Module is a system that employs user- or system-defined criteria to govern retrieval, ranking, and context assembly for adaptive evidence aggregation.
• It leverages mechanisms like conformal filtering and multi-query retrieval to ensure controlled evidence selection and measurable accuracy gains.
• Its modular design enables adaptive operations such as domain routing and targeted evidence selection, reducing hallucinations and improving decision usefulness.

A criterion-conditioned RAG (Retrieval-Augmented Generation) module refers to any modular operator, mechanism, or composite stage in a RAG pipeline whose activation and/or behavior is explicitly governed by user- or system-defined criteria. These criteria are used to shape retrieval, ranking, context assembly, or filtering phases according to coverage guarantees, domain constraints, evidence types, or downstream generator preferences. Criterion-conditioning enables RAG systems to move beyond monolithic, static retrieval, providing statistical guarantees, targeted evidence selection, or adaptive control over context construction—all in service of improved factuality, efficiency, or decision usefulness.

1. Formal Definition and Core Pattern

A criterion-conditioned module $M^{(c)}$ operates within a RAG pipeline as follows. Input is a tuple $I = (q, D, m)$, where $q$ is the input query, $D$ is a candidate document or snippet pool, and $m$ is optional auxiliary metadata (e.g., current generation confidence, query intent label). $M^{(c)}$ exposes:

• A functional interface $f: I \rightarrow O$ (e.g., filtering or reranking documents).
• A condition operator $g^{(c)}: I \rightarrow [0,1]$ or $g^{(c)}: I \rightarrow \{0,1\}$ (scored or Boolean).

$M^{(c)}$ is invoked only if $I = (q, D, m)$0 for some threshold $I = (q, D, m)$1 (scored) or $I = (q, D, m)$2 (Boolean). Thus, routing, scheduling, or processing are adaptively controlled via explicit criteria, yielding a modular, decoupled, and extensible architecture (Gao et al., 2024). Theoretical insights include:

• Decoupling “when to run” (via $I = (q, D, m)$3) from “what to do” (via $I = (q, D, m)$4) supports flexible orchestration.
• Probabilistic gating (softmax over criteria) generalizes hard routing, admitting uncertainty-aware ensembles.

Criteria used in practice include score thresholds (e.g., retrieval noise), domain-specific intent, working hypothesis state, or context size constraints.

2. Conformal Filtering and Coverage-Controlled Context Engineering

The conformal-prediction-based criterion-conditioned filter, as introduced in "Principled Context Engineering for RAG" (Chakraborty et al., 22 Nov 2025), inserts a Conformal Filter immediately after retrieval and before generation:

• Each retrieved snippet $I = (q, D, m)$5 receives a nonconformity score $I = (q, D, m)$6 (e.g., $I = (q, D, m)$7 cosine similarity or $I = (q, D, m)$8 LLM relevance rating).
• Using a calibration set of labeled pairs with $I = (q, D, m)$9, the $q$0-quantile threshold $q$1 is estimated:

$q$2

with $q$3 containing only positive calibration scores.

• Filtering then retains $q$4 if $q$5; with exchangeability, this ensures

$q$6

Empirical evaluations confirm tight control of relevant evidence retention (coverage), strong context compression (2–3× reduction), and—at strict or moderate filtering—stable or even improved downstream factual accuracy (e.g., ARGUE F1 improves by 0.03–0.04 for $q$7 on NeuCLIR) (Chakraborty et al., 22 Nov 2025). Filtering incurs no extra LLM training and is model-agnostic.

3. Decision-Useful and Hypothesis-Conditioned Criterion Control

"Criterion-conditioned" also applies to retrieval planning, as in Hypothesis-Conditioned Query Rewriting (HCQR) (Chang et al., 19 Mar 2026):

• Extraction of a working hypothesis $q$8: best-guess answer, discriminating features, confirming evidence, and rationale—performed by a prompt-based, zero-shot LLM.
• Construction of three criterion-specific queries:
  • $q$9: seeks evidence supporting $D$0.
  • $D$1: seeks passages distinguishing $D$2 from alternatives.
  • $D$3: targets verification of salient features.
• Each query issues independent retrieval; fused, deduplicated contexts are truncated to budget before use by the generator.

This targeted, intent-conditioned multi-query retrieval yields higher decision-useful evidence with a +5.9% (MedQA) and +3.6% (MMLU-Med) accuracy gain over single-query RAG. Ablation confirms each query role is additive, and the approach consistently improves the Decision-Useful Rate across benchmarks (Chang et al., 19 Mar 2026).

4. Preference- and Generator-Conditioned Criterion Modules

Criterion-conditioning is foundational in the Rank4Gen framework, which develops generator-aware document set selectors (Fan et al., 16 Jan 2026):

• The joint selection and ordering of subsets $D$4 are conditioned not just on query $D$5 but on the generator identity $D$6 (ID and textual profile) as a criterion.
• Training uses Direct Preference Optimization (DPO) to fit to preference pairs $D$7, labeled by improvement in response quality per generator as judged by an LLM.
• The input signature of the selector is $D$8, enabling tailored context assembly per downstream generator.

Rank4Gen outperforms or matches relevance-only baselines across diverse generators and tasks, with F1 gains up to +2 points and strong ablations showing clear benefit from criterion-aware supervision (Fan et al., 16 Jan 2026). The same architecture extends to task type, style constraints, or user preferences as criteria.

5. Knowledge-Driven and Adaptive Criterion Control

Structured knowledge graphs enable advanced criterion-conditioned control in Know³-RAG (Liu et al., 19 May 2025):

• An adaptive retrieval module conditions further retrieval on a reliability score $D$9—the sum of relative KGE deviation scores—computed over evidence triples extracted from the current answer and references.
• Only if $m$0 exceeds a dynamic threshold $m$1 is additional retrieval and context assembly triggered.
• Reference generation and filtering are also criterion-controlled, injecting KG entities and favoring semantically consistent and factually aligned contexts.

This multi-stage criterion-conditioning—across retrieval, generation, and filtering—substantially reduces hallucinations and improves answer reliability by synchronizing retrieval and filtering with evidence certifiability (Liu et al., 19 May 2025).

6. Orchestration, Routing, and Modular Design

The general pattern of criterion-conditioned modules underlies reconfigurable architectures as formalized in Modular RAG (Gao et al., 2024):

• Explicit router modules ($m$2) select among modules $m$3 based on criteria $m$4, admitting discrete, soft, or Boolean gating.
• Scheduling logic can repeat, branch, or fuse modules, enabling architectural patterns such as conditional, branching, and looping flows.
• Fusion operators (e.g., LLM-based summarization, ensemble weighting) integrate outputs from multiple activated modules.

This modularity supports use cases such as domain-routing (e.g., legal vs. medical), adaptive confidence-based iterative retrieval, and multi-modality processing—all grounded in explicit criterion evaluation (Gao et al., 2024).

7. Empirical Impact and Theoretical Guarantees

Empirical results across the literature demonstrate that criterion-conditioned modules yield:

• Tight coverage control, statistical guarantees on evidence retention, and context length reduction (Chakraborty et al., 22 Nov 2025).
• Substantial gains in decision-useful retrieval and QA accuracy under tight context budgets (Chang et al., 19 Mar 2026).
• Generator- and criterion-aligned evidence composition, increasing stability and response quality for varied LLM backends (Fan et al., 16 Jan 2026).
• Adaptive, knowledge-validated reference usage that demonstrably reduces hallucinations (Liu et al., 19 May 2025).

Theoretical grounds (as for conformal filters) guarantee criterion satisfaction under minimal distributional assumptions or statistical exchangeability. A plausible implication is that as more complex multi-criteria RAG systems are deployed, fine-grained modularity and criterion-conditionality will be necessary to optimize trade-offs in accuracy, relevance, and efficiency at scale.