Paper deep dive
Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models: From In-Context Prompting to Causal Retrieval-Augmented Generation
Prakhar Bansal, Shivangi Agarwal
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 4/10/2026, 2:07:13 AM
Summary
This survey provides a unified taxonomy and comparative analysis of contextual enrichment strategies for Large Language Models (LLMs), ranging from in-context prompting to Retrieval-Augmented Generation (RAG), GraphRAG, and CausalRAG. It evaluates these paradigms based on their ability to address knowledge, retrieval, and reasoning gaps, providing a deployment-oriented decision framework and a claim-audit synthesis to guide trustworthy NLP development.
Entities (5)
Relation Signals (4)
Prompting → suffersfrom → Lost-in-the-middle
confidence 95% · Long-context prompting alone suffers from placement sensitivity (lost-in-the-middle).
RAG → addresses → Knowledge gap
confidence 90% · RAG grounds generation in non-parametric external knowledge.
GraphRAG → improves → Relational reasoning
confidence 90% · GraphRAG improves multi-hop relational reasoning
CausalRAG → improves → Faithfulness
confidence 85% · Reported evidence from currently available benchmarks suggests improved answer faithfulness and retrieval precision over vanilla RAG
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models (LLMs) encode vast world knowledge in their parameters, yet they remain fundamentally limited by static knowledge, finite context windows, and weakly structured causal reasoning. This survey provides a unified account of augmentation strategies along a single axis: the degree of structured context supplied at inference time. We cover in-context learning and prompt engineering, Retrieval-Augmented Generation (RAG), GraphRAG, and CausalRAG. Beyond conceptual comparison, we provide a transparent literature-screening protocol, a claim-audit framework, and a structured cross-paper evidence synthesis that distinguishes higher-confidence findings from emerging results. The paper concludes with a deployment-oriented decision framework and concrete research priorities for trustworthy retrieval-augmented NLP.
Tags
Links
- Source: https://arxiv.org/abs/2604.03174v1
- Canonical: https://arxiv.org/abs/2604.03174v1
Trouble viewing inline? Open PDF directly →
Full Text
22,819 characters extracted from source content.
Expand or collapse full text
Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation Prakhar Bansal prakharb13@gmail.com Shivangi Agarwal shivangia@iiitd.ac.in Abstract Large language models (LLMs) encode vast world knowledge in their parameters, yet they remain fundamentally limited by static knowledge, finite context windows, and weakly structured causal reasoning. This survey provides a unified account of augmentation strategies along a single axis: the degree of structured context supplied at inference time. We cover in-context learning and prompt engineering, Retrieval-Augmented Generation (RAG), GraphRAG, and CausalRAG. Beyond conceptual comparison, we provide a transparent literature-screening protocol, a claim-audit framework, and a structured cross-paper evidence synthesis that distinguishes higher-confidence findings from emerging results. The paper concludes with a deployment-oriented decision framework and concrete research priorities for trustworthy retrieval-augmented NLP. Keywords: retrieval-augmented generation, causal reasoning, knowledge graphs, in-context learning, LLM augmentation, faithfulness, TrustNLP 1 Introduction Transformer-based language models have shown strong generality across language understanding and generation tasks. However, reliability in knowledge-intensive settings remains constrained by three recurring gaps: a knowledge gap (facts not encoded in model parameters), a retrieval gap (relevant evidence not surfaced), and a reasoning gap (causally incoherent outputs despite relevant evidence). We frame the progression from prompting to RAG, GraphRAG, and CausalRAG as a systematic response to these gaps. Each step introduces richer contextual structure and typically trades higher indexing complexity for better faithfulness and reasoning quality (Lewis et al., 2020; Edge et al., 2024; Wang et al., 2025). Contributions. This survey makes four contributions: (i) a unified contextual-enrichment taxonomy connecting prompting, RAG, GraphRAG, and CausalRAG; (i) a methodological protocol for literature selection and evidence grading; (i) a claim-audit table that links high-impact statements to primary sources and confidence levels; and (iv) a practitioner-focused decision framework for trustworthy deployment in high-stakes settings. Table 1: Three failure gaps in knowledge-intensive question answering and paradigm-level coverage (authors’ synthesis based on cited literature). Gap Prompting RAG GraphRAG CausalRAG Knowledge gap ∘ Yes Yes Yes Retrieval gap No ∘ Yes Yes Reasoning gap No No ∘ Yes Scope and Organization. This survey targets ML practitioners and researchers seeking a coherent comparison of retrieval augmentation paradigms. We do not propose a new algorithm. Instead, we formalize and compare established methods using common notation and evaluation criteria. 2 Survey Methodology 2.1 Literature Collection Protocol We conducted a structured literature collection across peer-reviewed venues and major preprint repositories for retrieval-augmented generation and causal reasoning in LLM systems. The search window for this draft ends in March 2026. Candidate papers were screened by title and abstract, then full-text reviewed for methodological relevance. Inclusion Criteria. Papers were included if they: (i) propose, evaluate, or systematically analyze prompting, RAG, GraphRAG, or CausalRAG pipelines; (i) report concrete retrieval or generation behavior rather than purely conceptual commentary; and (i) provide enough methodological detail to support comparative interpretation. Exclusion Criteria. We excluded works lacking technical detail, duplicate versions without substantive updates, and papers that do not expose evidence supporting claims about faithfulness, retrieval quality, or causal reasoning. 2.2 Evidence Grading We use three evidence grades in this survey. High confidence indicates repeated support across multiple studies and settings. Medium confidence indicates strong but still limited or heterogeneous evidence. Emerging evidence indicates promising results with narrow benchmark or dataset coverage. These grades guide wording throughout the manuscript. 3 Context in Large Language Models 3.1 What is Context? In transformer architectures, context is the visible token sequence x=(x1,…,xn)x=(x_1,…,x_n) during generation. We distinguish three context types: parametric context (stored in model weights), in-context knowledge (prompt-provided information), and retrieved context (externally retrieved evidence inserted at inference time). In-Context Learning and Prompt Engineering. Few-shot prompting demonstrates that large models can adapt to novel tasks from examples without gradient updates (Brown et al., 2020). Prompt engineering broadens this paradigm with zero-shot prompting, chain-of-thought prompting, and retrieval-augmented prompting (Liu et al., 2023; Wei et al., 2022). Context-Window Limitations. Even with long-context models, lost-in-the-middle degradation and noise from irrelevant inserted text remain practical failure modes (Liu et al., 2024). Separately, purely parametric knowledge can become stale as world facts change, which motivates retrieval-grounded generation (Petroni et al., 2019; Lewis et al., 2020). 4 Retrieval-Augmented Generation (RAG) 4.1 Background and Definition RAG grounds generation in non-parametric external knowledge. Let D=d1,…,dmD=\d_1,…,d_m\ be a corpus indexed into chunk embeddings. For query q, the retriever selects top-k chunks; the generator conditions on q plus retrieved chunks (Lewis et al., 2020). Retrieval Architectures. Sparse retrieval (e.g., BM25) is interpretable and efficient but can miss semantic matches (Robertson and Zaragoza, 2009). Dense retrieval captures semantic similarity but is costlier and domain-sensitive (Karpukhin et al., 2020). Hybrid pipelines often improve robustness by combining both (Chen et al., 2023). Advanced Variants. Recent variants include iterative or multi-hop RAG, Self-RAG (adaptive retrieval and reflection), and modular RAG pipelines with query rewriting, reranking, and post-filtering (Trivedi et al., 2023; Asai et al., 2024; Gao et al., 2023). 4.2 Limitations Vanilla RAG is limited by contextual fragmentation, semantic-similarity bias, and limited global summarization for corpus-level questions. 5 Graph Retrieval-Augmented Generation (GraphRAG) 5.1 Motivation and Formulation GraphRAG replaces flat chunk retrieval with knowledge-graph indexing. During indexing, systems extract triples (head,relation,tail)(head,relation,tail) and often produce community summaries via graph clustering. At query time, local or global graph retrieval provides structured evidence to generation (Edge et al., 2024; Peng et al., 2024; Han et al., 2025). Advantages. GraphRAG improves multi-hop relational reasoning, supports global thematic synthesis, and improves traceability of generated claims. Limitations. Key bottlenecks include expensive graph construction, entity resolution noise, temporal rigidity, and associative (rather than explicitly causal) edges. 6 Causal Retrieval-Augmented Generation (CausalRAG) 6.1 Motivation Causal questions (“Why did X happen?”, “What if Y changed?”) require more than topical retrieval. CausalRAG introduces directed causal structure into retrieval to support faithful, interpretable causal explanations (Pearl, 2009; Samarajeewa et al., 2024; Wang et al., 2025). Causal Graph Foundation. Using structural causal modeling intuition, causal graphs encode directed relations and support observational, interventional, and counterfactual reasoning modes. Pipeline. CausalRAG extracts directed causal triples from source text, indexes them, and retrieves causally connected subgraphs via graph walks from query-linked seed nodes. Generation is then conditioned on a causally grounded evidence narrative. Empirical Pattern. Reported evidence from currently available benchmarks suggests improved answer faithfulness and retrieval precision over vanilla RAG, with stronger causal coherence than purely associative graph retrieval (Wang et al., 2025). Because benchmark settings vary across papers, these results should be interpreted as indicative rather than directly comparable. 7 Unified Comparative Framework 7.1 Contextual Enrichment Axis Prompting, RAG, GraphRAG, and CausalRAG can be viewed as a monotonic increase in context structure. This yields a practical trade-off: higher indexing complexity in exchange for higher reasoning faithfulness. 7.2 Quantitative Evidence Across Four Paradigms Table 2 summarizes representative quantitative results from primary papers across the four paradigms. Each row states the exact systems being compared, the metric and unit, and reported values with deltas. Cross-row comparisons across different papers should not be treated as head-to-head performance rankings. Table 2: Quantitative evidence matrix with explicit metric definitions and scope constraints. All values are reported in the cited source studies. Paradigm Source Dataset Systems compared Metric (unit) Reported values Delta Prompting (CoT) Wei et al. (2022) GSM8K PaLM 540B: CoT vs standard prompting Solve rate (%) 57.0 vs 18.0 +39.0 pts RAG Lewis et al. (2020) Natural Questions (Open-domain QA) RAG vs T5-11B (closed-book) Exact Match (EM, %) 44.5 vs 34.5 +10.0 pts GraphRAG-Global (slice comparison) Wang et al. (2025) OpenAlex single-paper case study GraphRAG-Global: abstract vs full text Composite score (author-defined) 49.98 vs 76.37 +26.39 GraphRAG-Local (slice comparison) Wang et al. (2025) OpenAlex single-paper case study GraphRAG-Local: abstract vs full text Composite score (author-defined) 62.29 vs 84.27 +21.98 CausalRAG (slice comparison) Wang et al. (2025) OpenAlex single-paper case study CausalRAG: abstract, intro, and full text Composite score (author-defined) 72.43, 74.86, 91.69 +19.26 (abstract to full) CausalRAG (k=s sweep) Wang et al. (2025) OpenAlex CausalRAG: k=s=1 vs k=s=5 Author-reported score 0.534 vs 0.824 +0.290 Interpretation note: All numerical deltas in Table 2 are within-study comparisons under each cited paper’s own models, dataset split, and evaluation protocol; they are not cross-paper rankings. Metric note: “Composite score (author-defined)” follows Wang et al. (2025), which reports an aggregate score over answer faithfulness, context precision, and context recall (Figure 5). The paper does not provide an explicit weighting formula for this aggregate in the main text. Scope note: The GraphRAG and CausalRAG slice rows from Wang et al. (2025) come from a controlled follow-up case study on one marketing paper and one question, evaluated on abstract/introduction/full-text slices; this is not a cross-dataset leaderboard. Table 3: Comparative summary across paradigms (authors’ synthesis). Criterion Prompting RAG GraphRAG CausalRAG Indexing cost None Low to Medium High High to Very High Relational reasoning Limited Limited Strong Strong Causal reasoning Weak Weak Partial Strong Global summary Weak Moderate Strong Strong Typical faithfulness Low to Medium Medium Medium to High High When to Use Which. Prompting is suitable for lightweight, in-window tasks. RAG is a practical default for factual grounding at scale. GraphRAG is preferred for multi-hop/entity-centric questions and corpus-level synthesis. CausalRAG is best for root-cause and high-stakes explanatory tasks where interpretability is critical. Worked Example. For an industry finance query such as “Why did operating margin decline in Q3 after the pricing update?”, prompting can produce plausible but ungrounded narratives. RAG can retrieve earnings-call passages and management commentary. Graph-based retrieval can connect entities such as product line, region, and cost center across filings. CausalRAG can prioritize directed causal chains—for example, price reduction → lower unit revenue → margin compression—while keeping each step linked to source evidence. 7.3 Claim Audit Table Table 4 documents high-impact technical claims, their supporting citations, and a confidence label that reflects evidence maturity in the literature. Table 4: Claim audit for key statements in this survey (authors’ evidence grading with explicit source attribution). Claim Primary support Confidence Notes RAG improves factual grounding versus prompting-only baselines in knowledge-intensive QA. Lewis et al. (2020); Chen et al. (2023) High Established across multiple benchmarks. CausalRAG improves faithfulness and causal coherence over vanilla RAG and associative graph retrieval in reported settings. Wang et al. (2025); Samarajeewa et al. (2024) Medium Evidence is promising but still limited in scale. GraphRAG supports stronger relational and global-corpus reasoning than flat top-k chunk retrieval. Edge et al. (2024); Han et al. (2025); Peng et al. (2024) Medium to High Strong conceptual fit; evaluation setups vary. Long-context prompting alone suffers from placement sensitivity (lost-in-the-middle). Liu et al. (2024) High Strong evidence in controlled long-context evaluations. Automated causal extraction remains error-prone and expensive in current LLM pipelines. Jiralerspong et al. (2024); Wang et al. (2025) Medium Active research area; methods evolving quickly. 8 Limitations Open challenges include scalable causal extraction, standardized causal evaluation metrics, dynamic maintenance of evolving causal graphs, multilingual causal retrieval, and integration with agentic LLM workflows. Threats to Validity. Three threats shape the interpretation of this survey. First, cross-paper experimental heterogeneity limits strict numerical comparability. Second, publication and survivorship bias may over-represent positive outcomes. Third, recency effects are substantial in fast-moving LLM research; conclusions should be periodically re-audited as new evaluations appear. 9 Broader Impact This survey discusses trustworthiness-oriented retrieval augmentation strategies and is intended to support safer, evidence-grounded use of LLMs in high-stakes domains. Potential risks include overconfidence in automatically extracted causal structure, inappropriate transfer of benchmark findings to deployment settings, and misuse for persuasive but weakly grounded narratives. We mitigate these risks by explicitly labeling evidence confidence, separating within-study findings from cross-study interpretation, and highlighting unresolved limitations and evaluation gaps. 10 Alignment with TrustNLP This survey aligns with TrustNLP themes on causal inference, faithfulness and safety, hallucination reduction, interpretability, and trustworthy high-stakes applications. Workshop Relevance. The paper emphasizes trustworthiness through evidence-aware claim calibration, explicit uncertainty reporting, and a deployment-oriented view of failure modes. This framing is designed to be directly useful for TrustNLP reviewers evaluating methodological rigor and practical trustworthiness impact. 11 Conclusion Context enrichment for LLMs is best treated as an engineering continuum rather than a binary choice. CausalRAG is a promising direction for faithful causal question answering, but it carries substantial indexing and maintenance costs and is supported by a still-emerging evidence base. Future progress depends on scalable causal extraction, robust evaluation standards, and tighter integration with planning-capable LLM agents. References A. Asai, Z. Wu, Y. Wang, A. Sil, and H. Hajishirzi (2024) Self-RAG: learning to retrieve, generate, and critique through self-reflection. In International Conference on Learning Representations (ICLR), Cited by: §4.1. T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei (2020) Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §3.1. J. Chen, H. Lin, X. Han, and L. Sun (2023) Benchmarking large language models in retrieval-augmented generation. arXiv preprint arXiv:2309.01431. Cited by: §4.1, Table 4. D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, and J. Larson (2024) From local to global: a graph RAG approach to query-focused summarization. arXiv preprint arXiv:2404.16130. Cited by: §1, §5.1, Table 4. S. Es, J. James, L. Espinosa-Anke, and S. Schockaert (2023) RAGAS: automated evaluation of retrieval-augmented generation. arXiv preprint arXiv:2309.15217. Cited by: Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation. Y. Gao, Y. Xiong, X. Gao, K. Jia, J. Pan, Y. Bi, Y. Dai, J. Sun, M. Wang, and H. Wang (2023) Retrieval-augmented generation for large language models: a survey. arXiv preprint arXiv:2312.10997. External Links: 2312.10997, Document, Link Cited by: §4.1. Z. Guo, L. Xia, Y. Yu, T. Ao, and C. Huang (2024) LightRAG: simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779. Cited by: Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation. H. Han, Y. Wang, H. Shomer, K. Guo, J. Ding, Y. Lei, M. Halappanavar, R. A. Rossi, S. Mukherjee, X. Tang, Q. He, Z. Hua, B. Long, T. Zhao, N. Shah, A. Javari, Y. Xia, and J. Tang (2025) Retrieval-augmented generation with graphs (GraphRAG). arXiv preprint arXiv:2501.00309. Cited by: §5.1, Table 4. T. Jiralerspong, X. Chen, Y. More, V. Shah, and Y. Bengio (2024) Efficient causal graph discovery using large language models. arXiv preprint arXiv:2402.01207. Cited by: Table 4, Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation. V. Karpukhin, B. Oguz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W. Yih (2020) Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), Cited by: §4.1. P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. K"uttler, M. Lewis, W. Yih, T. Rockt"aschel, S. Riedel, and D. Kiela (2020) Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems (NeurIPS), p. 9459–9474. Cited by: §1, §3.1, §4.1, Table 2, Table 4. N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang (2024) Lost in the middle: how language models use long contexts. Transactions of the Association for Computational Linguistics 12, p. 157–173. Cited by: §3.1, Table 4. P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neubig (2023) Pre-train, prompt, and predict: a systematic survey of prompting methods in natural language processing. ACM Computing Surveys. Cited by: §3.1. J. Pearl (2009) Causality: models, reasoning, and inference. Cambridge University Press. Cited by: §6.1. B. Peng, Y. Zhu, Y. Liu, X. Bo, H. Shi, C. Hong, Y. Zhang, and S. Tang (2024) Graph retrieval-augmented generation: a survey. arXiv preprint arXiv:2408.08921. Cited by: §5.1, Table 4. F. Petroni, T. Rockt"aschel, S. Riedel, P. Lewis, A. Bakhtin, Y. Wu, and A. Miller (2019) Language models as knowledge bases?. In Proceedings of EMNLP-IJCNLP, Cited by: §3.1, Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation. S. Robertson and H. Zaragoza (2009) The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval 3 (4), p. 333–389. Cited by: §4.1. C. Samarajeewa, D. D. Silva, E. Osipov, D. Alahakoon, and M. Manic (2024) Causal reasoning in large language models using causal graph retrieval augmented generation. In 2024 16th International Conference on Human System Interaction (HSI), p. 1–6. External Links: Document Cited by: §6.1, Table 4. B. Sch"olkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y. Bengio (2021) Toward causal representation learning. Proceedings of the IEEE 109 (5), p. 612–634. Cited by: Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation. H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal (2023) Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Toronto, Canada, p. 10014–10037. External Links: Document, Link Cited by: §4.1. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin (2017) Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation. N. Wang, X. Han, J. Singh, J. Ma, and V. Chaudhary (2025) CausalRAG: integrating causal graphs into retrieval-augmented generation. In Findings of the Association for Computational Linguistics: ACL 2025, p. 22680–22693. External Links: Document, Link Cited by: §1, §6.1, §6.1, §7.2, §7.2, Table 2, Table 2, Table 2, Table 2, Table 4, Table 4. J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V. Le, and D. Zhou (2022) Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §3.1, Table 2. S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao (2023) ReAct: synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), Cited by: Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation. P. Zhao, H. Zhang, Q. Yu, Z. Wang, Y. Geng, F. Fu, L. Yang, W. Zhang, J. Jiang, and B. Cui (2025) Retrieval-augmented generation for AI-generated content: a survey. Data Science and Engineering. External Links: Document Cited by: Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models From In-Context Prompting to Causal Retrieval-Augmented Generation.