Beyond the Parameters: A Technical Survey of Contextual Enrichment in Large Language Models: From In-Context Prompting to Causal Retrieval-Augmented Generation
| Authors | Prakhar Bansal & Shivangi Agarwal |
| Year | 2026 |
| Field | NLP |
| arXiv | 2604.03174 |
| Download | |
| Categories | cs.CL, cs.AI |
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.
Engineering Breakdown
Plain English
This paper is a comprehensive survey that categorizes and compares different strategies for augmenting large language models with structured context at inference time, ranging from simple prompt engineering to sophisticated graph-based retrieval systems. The authors organize these approaches along a single axis—the degree of structured context provided—covering in-context learning, RAG, GraphRAG, and CausalRAG. Beyond conceptual comparison, they contribute a transparent literature-screening protocol, a claim-audit framework, and cross-paper evidence synthesis to separate high-confidence findings from emerging results. The paper culminates in a practical decision framework for deployment and identifies concrete research priorities for trustworthy augmentation systems.
Core Technical Contribution
The core contribution is a unified taxonomy that organizes knowledge augmentation strategies on a single interpretable dimension: the amount and structure of context supplied at inference time. Rather than treating RAG, GraphRAG, and CausalRAG as disconnected approaches, the authors reveal them as points on a spectrum from unstructured retrieval to richly structured causal graphs. A secondary but important contribution is the development of a claim-audit framework and transparent literature-screening protocol that provides a rigorous methodology for synthesizing evidence across papers—addressing a real problem in ML survey work where claims are often repeated without verification. This methodological contribution allows practitioners to distinguish between empirically validated findings and speculative results with explicit confidence levels.
How It Works
The survey organizes LLM augmentation strategies hierarchically, with the primary axis being the structure and organization of context provided at inference. At one end sits in-context learning and prompt engineering, which supply unstructured text examples or instructions that the model must parse and apply implicitly. Moving along the spectrum, Retrieval-Augmented Generation (RAG) explicitly retrieves relevant documents from a corpus and concatenates them with the query, providing semi-structured context. GraphRAG extends this by organizing retrieved information as a knowledge graph where entities and relations are explicitly represented, allowing the model to reason over structured relationships rather than flat text. CausalRAG further refines this by explicitly modeling causal relationships between concepts, enabling the system to reason about interventions and counterfactuals. The evidence synthesis methodology involves screening papers against explicit inclusion criteria, extracting claims with their supporting evidence, auditing claims against source material, and rating confidence levels based on the quantity and quality of empirical support across papers.
Production Impact
For teams building production systems, this paper provides a practical decision framework for choosing augmentation strategies based on inference latency budgets, retrieval quality requirements, and reasoning complexity. Simple in-context learning requires no external data infrastructure but sacrifices reasoning depth; RAG adds retrieval latency (typically 100-500ms) but dramatically improves factuality; GraphRAG further increases complexity but enables multi-hop reasoning over interconnected facts; CausalRAG adds causal inference capabilities at the cost of explicit causal graph construction and maintenance. The structured taxonomy helps engineers reason through trade-offs: if your application needs real-time responses (sub-100ms), in-context learning or lightweight RAG is forced; if you have domain-specific reasoning requirements, building a knowledge graph becomes justified. The claim-audit framework is particularly valuable for validating vendor claims about RAG effectiveness—many papers overstate improvements or don't control for confounding variables like larger context windows, and this framework provides a methodology to catch those inflations.
Limitations and When Not to Use This
The paper is a survey that synthesizes existing work rather than proposing novel algorithms, so it doesn't solve the fundamental technical challenges in knowledge augmentation—like how to rank retrieved documents without degrading reasoning, how to scale knowledge graph construction to billions of facts, or how to verify causal claims in learned representations. The framework assumes that augmentation strategies can be cleanly ordered along a single spectrum (structured context), but real-world effectiveness often depends on task-specific factors like domain shift, question complexity, and the quality of the underlying retrieval index, which aren't fully captured by this one-dimensional view. The evidence synthesis methodology, while rigorous, is only as good as the papers it reviews; emerging techniques not yet published in major venues won't be represented, and the confidence levels depend on reviewers' interpretations of evidence quality. Finally, the paper doesn't address computational costs in detail—building and maintaining GraphRAG or CausalRAG systems requires significant engineering effort to construct knowledge graphs and run graph algorithms, costs that may outweigh benefits for many applications.
Research Context
This work builds directly on the RAG literature (Lewis et al. 2020, and subsequent refinements) and extends it to structured knowledge representation strategies like knowledge graphs and causal models. It positions itself within the broader context of LLM augmentation research that emerged around 2023-2024 as practitioners discovered that LLMs alone were insufficient for tasks requiring long-horizon reasoning, factual grounding, or domain-specific knowledge. The survey is timely because the field has fragmented into many specialized augmentation approaches (GraphRAG, HyDE, Self-RAG, etc.) without a unifying framework, leading to confusion about which approach to use when. By providing a principled taxonomy and evidence synthesis methodology, the paper opens research directions in: (1) developing metrics that can fairly compare augmentation strategies across domains, (2) understanding when structured vs. unstructured context provides greater benefit, and (3) building systems that adaptively choose augmentation strategies based on query complexity.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::