retrieval augmented generation: one bounded, context-dependent signal across receipts

Source literature boundary memo

Research question

Does retrieval augmented generation show a consistent direction-bearing association in the selected source bundle, and where do null/mixed or context-only receipts bound the claim?

Selection criteria

The source-literature selector kept retrieval augmented generation because the candidate bundle met the public source rule: 5 citable papers, 5 distinct fact-backed source identities, topic-overlapping source facts, and enough shared scope to compare metric/context disagreement. It excludes duplicate reports, metadata-only title matches, off-topic papers, and sources without fact-level extraction before treating the bundle as a coherent scoping front rather than proof of a policy or market conclusion.

Plain-language synthesis

3 of 5 selected receipts are direction-bearing for the selected source contexts; 0 receipt(s) are null/mixed and 2 are context/model only. This is a bounded source-literature signal, not a pooled effect.

Boundary map

• A Retrieval-Augmented Generation Framework for Traditional Chinese Medicine Herb Recommendation Using Symptom-Focused and Ingredient-Based Embeddings [primary; 2026] doi:10.65205/jcct.2026.e3516
  • Bounded source claim: The baseline LLM demonstrated strong performance across multiple metrics, including accuracy (0.1900) and NDCG@5 (0.1475), reflecting substantial pre-trained medical knowledge.
  • Claim bounds: setting=rag accuracy tasks; exposure=Retrieval-Augmented Generation Framework; comparator/reference=LLM demonstrated strong performance across multiple metrics, including accuracy (0.1900)
  • Effect accounting: descriptive/modeling context only; this receipt does not test an effect of retrieval augmented generation on a performance endpoint.
  • Population/setting: rag accuracy tasks
  • Policy/exposure/practice: Retrieval-Augmented Generation Framework
  • Comparator/reference: LLM demonstrated strong performance across multiple metrics, including accuracy (0.1900)
• Evaluating Retrieval-Augmented Generation Variants for Natural Language-Based SQL and API Call Generation [primary; 2026] doi:10.48550/arxiv.2602.07086
  • Bounded source claim: Critically, CoRAG proves most robust in hybrid documentation settings, achieving statistically significant improvements in the combined task (10.29% exact match vs. 7.45% for standard RAG), driven primarily by superior SQL generation performance (15.32% vs. 11.56%).
  • Claim bounds: setting=combined; exposure=RAG; comparator/reference=7.45% for standard RAG), driven primarily by superior SQL generation performance (15.32%
  • Population/setting: combined
  • Policy/exposure/practice: RAG
  • Comparator/reference: 7.45% for standard RAG), driven primarily by superior SQL generation performance (15.32%
• A retrieval-augmented generation large language model framework for accurate dementia identification from electronic health records [primary; 2026] doi:10.64898/2026.01.24.26344477
  • Bounded source claim: ResultsThe RAG-based classifier achieved the highest performance (F1=0.933, sensitivity=91.1%, PPV=95.5%) compared to rule-based (F1=0.823, sensitivity=81.1%, PPV=83.5%) and keyword-filtered LLM (F1=0.903, sensitivity=91.7%, PPV=88.6%).
  • Claim bounds: setting=rag F1 tasks; exposure=RAG; comparator/reference=rule-based (F1=0.823, sensitivity=81.1%, PPV=83.5%) and keyword-filtered LLM (F1=0.903, s
  • Effect accounting: descriptive/modeling context only; this receipt does not test an effect of retrieval augmented generation on a performance endpoint.
  • Population/setting: rag F1 tasks
  • Policy/exposure/practice: RAG
  • Comparator/reference: rule-based (F1=0.823, sensitivity=81.1%, PPV=83.5%) and keyword-filtered LLM (F1=0.903, s
• Integrating Dense, Sparse, and Graph-Based Approaches in Financial Data Analysis for a Retrieval-Augmented Generation Framework [primary; 2026] doi:10.1109/acdsa67686.2026.11467963
  • Bounded source claim: Results show that integrating a graph-based retriever improved context recall by 63%, answer correctness by 31%, and overall performance by 12% compared to flattened text retrieval.
  • Claim bounds: setting=rag recall tasks; exposure=Integrating Dense, Sparse, and Graph-Based Approaches; comparator/reference=flattened text retrieval
  • Population/setting: rag recall tasks
  • Policy/exposure/practice: Integrating Dense, Sparse, and Graph-Based Approaches
  • Comparator/reference: flattened text retrieval
• Improving Retrieval-Augmented Generation Performance Using the MAF-RAG Architecture, EVR–VOR Vector Retrieval, and Multi-Agent Fallback Reasoning [primary; 2026] doi:10.30871/jaic.v10i1.11738
  • Bounded source claim: The results show that the proposed MAF-RAG significantly outperforms the baseline system, achieving a mean F1-score of 0.556, an improvement of 18.8% over the Enhanced Baseline (mean F1-score = 0.469) and a 70.0% improvement over the Legacy Baseline (mean F1-score = 0.327).
  • Claim bounds: setting=rag F1 tasks; exposure=RAG; comparator/reference=the baseline system
  • Population/setting: rag F1 tasks
  • Policy/exposure/practice: RAG
  • Comparator/reference: the baseline system

Source synthesis

Bounded signal: retrieval augmented generation is only a source-level context map; the selected receipts do not establish one pooled effect.

This receipt-backed scoping note has one bounded signal: retrieval augmented generation shows policy/exposure estimates plus separate descriptive evidence across this 5-source primary bundle (2026-2026). Evidence role grouping: direction-bearing receipts: 3; null/mixed metric-scope caveat receipts: 0; context/antecedent/model receipts: 2 excluded from effect support. The source facts cover 4 population/setting context(s) and 3 policy/exposure/practice context(s), so this is a scoping signal about where settings/designs diverge, without establishing a causal, policy-prescriptive, market-generalized, or pooled econometric claim. Population/setting counts are context descriptors only; they are not weighting, pooling, or aggregation evidence. The listed estimates remain source-specific across metrics and settings; they are not pooled or averaged. This is a separated policy/setting map, not a unified pooled economics claim. Named setting scope includes combined, rag F1 tasks, rag accuracy tasks, and rag recall tasks. Within-vs-across outcome rule: direction-bearing rows are only compared within the selected source contexts; unrelated receipt families are not treated as one outcome. Concrete contrast: directional association: Evaluating Retrieval-Augmented Generation Variants for Natural Language-Based SQL and API Call Generation: Critically, CoRAG proves most robust in hybrid documentation settings, achieving statistically significant...; descriptive/modeling: A Retrieval-Augmented Generation Framework for Traditional Chinese Medicine Herb Recommendation Using Symptom-Focused and Ingredient-Based Embeddings: The baseline LLM demonstrated strong performance across multiple metrics, including accuracy (0.1900) and....

Role definitions: direction-bearing rows carry metric-specific effect or association text; null/mixed rows carry rejected or non-convergent metric evidence; context/model rows rank, model, or contextualize adjacent constructs. Interpretation: keep these rows separate; do not pool them or treat antecedent/modeling rows as the same estimand.

Evidence matrix

Matrix guard: effect-bearing rows below are metric-specific source facts, not a pooled comparison; context-only rows are excluded from effect support.

Effect-bearing comparison

Outcome family	Receipt	Evidence role	Population/setting	Metric	Extracted finding
outcome-specific	Evaluating Retrieval-Augmented Generation Variants for Natural...	directional association	combined	-	Critically, CoRAG proves most robust in hybrid documentation settings, achieving statistically significant...
outcome-specific	Integrating Dense, Sparse, and Graph-Based Approaches in Financial Data...	directional association	rag recall tasks	-	Results show that integrating a graph-based retriever improved context recall by 63%, answer correctness by...
outcome-specific	Improving Retrieval-Augmented Generation Performance Using the MAF-RAG...	directional association	rag F1 tasks	-	The results show that the proposed MAF-RAG significantly outperforms the baseline system, achieving a mean...

Context-only receipts

Outcome family	Receipt	Evidence role	Population/setting	Metric	Extracted finding
modeling-context	A Retrieval-Augmented Generation Framework for Traditional Chinese...	descriptive/modeling	rag accuracy tasks	-	The baseline LLM demonstrated strong performance across multiple metrics, including accuracy (0.1900) and...
modeling-context	A retrieval-augmented generation large language model framework for...	descriptive/modeling	rag F1 tasks	-	ResultsThe RAG-based classifier achieved the highest performance (F1=0.933, sensitivity=91.1%, PPV=95.5%)...

Audit note: effect-bearing rows stay metric-specific; context-only rows are excluded from effect support; role counts below keep direction-bearing, null/mixed metric-scope caveat, and context-only receipts separate.

Evidence role definitions

• directional association: source-level direction with design caveat; retrieval_augmented_generation is the policy, exposure, method, or practice linked to the named metric, not a pooled effect-size estimate or efficacy verdict.
• descriptive/modeling: the receipt reports modelling or prediction rather than a policy-effect estimate.

Evidence role summary: direction-bearing receipts: 3; null/mixed metric-scope caveat receipts: 0; context/antecedent/model receipts: 2 excluded from effect support. Direction labels for audit: descriptive/modeling: 2 receipt(s) | directional association: 3 receipt(s).

Specific moderators in this bundle are population/indication (combined; rag F1 tasks; rag accuracy tasks; rag recall tasks), study design/evidence type (primary).

Context separation

Population/settings are separated as receipt context: combined, rag F1 tasks, rag accuracy tasks, and rag recall tasks. The selected receipts group because each carries a fact-level extraction for retrieval augmented generation; they separate by context (other source context) and metric, so they are not interchangeable evidence for one pooled claim.

Boundary limits

Source-literature boundary for retrieval augmented generation: the listed sources define one bounded, context-dependent signal across separate source contexts. This memo does not claim causality, policy prescription, a pooled elasticity estimate, or a market-generalized effect across the sources. Material limitations: small 5-source bundle; no pooled estimate is possible; outlet/tier heterogeneity is scope, not weight; method/model receipts without direct effect estimates are context only; outcomes are not harmonized across studies. The signal is purely descriptive of source-level direction and scope; it cannot support a causal, policy-prescriptive, or pooled elasticity inference, and pooling across these designs would be inappropriate. Effect-support accounting: 2 of 5 receipt(s) is context/modeling-only and contributes no effect estimate; 3 receipt(s) are direction-bearing and 0 receipt(s) are null/mixed metric-scope caveats.

What would weaken this

• This scoping signal would weaken if the null/mixed metric replicates in matched designs, if direction-bearing rows fail to reproduce within their named metric family, or if context/model rows become the only topic-overlapping receipts.

Next gaps

A stronger memo needs one matched design: one setting, one policy/exposure, one comparator/reference group, and one named metric. If retrieval augmented generation is promoted beyond a scoping note, the next run should select sources sharing one context family rather than spanning other source context.