From Search to Reasoning:
A Five-Level RAG Capability Framework for Enterprise Data

Abstract

Retrieval-Augmented Generation (RAG) has emerged as the standard paradigm for answering questions on enterprise data. Traditionally, RAG has centered on text-based semantic search and re-ranking. However, this approach falls short when dealing with questions beyond data summarization or non-text data. This has led to various attempts to supplement RAG to bridge the gap between RAG, the implementation paradigm, and the question answering problem that enterprise users expect it to solve.

Given that contemporary RAG is a collection of techniques rather than a defined implementation, discussion of RAG and related question-answering systems benefits from a problem-oriented understanding.

We propose a new classification framework (L1–L5) to categorize systems based on data modalities and task complexity of the underlying question answering problems: L1 (Surface Knowledge of Unstructured Data) through L4 (Reflective and Reasoned Knowledge) and the aspirational L5 (General Intelligence). We also introduce benchmarks aligned with these levels and evaluate four state-of-the-art platforms: LangChain, Azure AI Search, OpenAI, and Corvic AI. Our experiments highlight the value of multi-space retrieval and dynamic orchestration for enabling L1–L4 capabilities. We empirically validate our findings using diverse datasets indicative of enterprise use cases.

1 Introduction

Large Language Models (LLMs) have introduced a new paradigm (Brown et al.,, 2020) in data processing and utilization, and in enterprise environments, where private data represents a significant untapped asset, LLMs promise to transform vast stores of yet unexploited data into actionable insights. The potential applications are wide-ranging: from enhancing automated customer interactions to facilitating sophisticated decision-making processes. And, by leveraging a deeper understanding of data through LLMs, enterprises can augment their workforce with AI automation, thereby boosting productivity and operational efficiency.

However, the use of LLMs in enterprise scenarios has challenges. One significant hurdle is the tendency of LLMs to generate “hallucinations” or content-fabricated information not present in the input data (Huang et al.,, 2025). This issue of hallucination, coupled with data incompleteness, can lead to unreliable outputs that jeopardize business decisions with potentially catastrophic consequences. As enterprises rely heavily on data integrity, the prevalence of misinformation generated by LLMs presents an existential risk.

To address these challenges, recent focus has shifted towards mitigating the hallucination phenomenon. Retrieval-Augmented Generation (RAG) (Lewis et al.,, 2020) has emerged as a promising approach, enhancing LLMs by integrating them with contextually relevant, enterprise-specific data (Izacard and Grave,, 2021; Karpukhin et al.,, 2020). This integration aims to ground an LLM’s responses with known facts and reduce the probability of fabrication (Shuster et al.,, 2021). Yet, despite its potential, RAG still grapples with the same challenges as using LLMs without RAG (Gao et al.,, 2024), such as maintaining response accuracy and relevance. This paper studies these issues, exploring innovative strategies to harness the full potential of LLMs in enterprise settings while ensuring accuracy and reliability crucial for business applications.

Figure 1: Generic RAG pipeline: Architecture for L1 (surface knowledge of unstructured data) pipeline.

Figure 2: Corvic AI: Mixture of Spaces (MoS) and Adaptive Chain of Actions (ACoA) Architecture for L4 (reflective and reasoned knowledge) pipeline.

2 Background

Retrieval-Augmented Generation (RAG) (Fig. 1) has emerged as a standard way to address the challenges of hallucination and incomplete responses of Large Language Models (LLMs) in data-sensitive environments. It adds a retrieval mechanism that fetches contextually relevant information from a private datasets. The original query is then augmented with this additional context before being passed to an LLM for question-answering (Ram et al.,, 2023). This method not only curtails the generation of hallucinated content but also enhances the completeness and factual accuracy of the responses (Lewis et al.,, 2020) with respect to the private data. Despite its advancements, the RAG approach is still not a panacea. For example, Magesh et al., (2024) have shown significant hallucination and incompleteness for legal questions, noting hallucination and incompleteness rates across systems between 10% and 60%. These findings have been observed across various other problem domains (Amugongo et al.,, 2025).

One limitation of RAG is that retrieval is through an embedding space that only captures language understanding and semantic similarity (Xiong et al.,, 2021). While useful for certain kinds of problems, this single-faceted approach falls short when dealing with complex queries that require nuanced understanding and multi-faceted contextual grounding. In an attempt to address this issue, the majority of the research in RAG systems has been focused on re-ranking algorithms (Nogueira and Cho,, 2020; Nogueira et al.,, 2019; Zhuang and Zuccon,, 2021), which refine the selection of retrieved documents based on their perceived relevance to the input query (Guu et al.,, 2020; Karpukhin et al.,, 2020).

However, the efficacy of re-ranking is inherently limited by the quality of the initial retrieval phase. If the retrieval component fails to bring forward the relevant context, even the most sophisticated re-ranker cannot compensate for the initial deficiency. This decoupling of retrieval and re-ranking phases often leads to an accuracy bottleneck (Yu et al.,, 2023), which highlights the need for a different approach. Enhancing the retrieval mechanisms to consider multiple dimensions of semantic relevance and context variability (Khattab and Zaharia,, 2020; Santhanam et al.,, 2022) paves the way for more robust and reliable outputs from LLMs, particularly in complex and dynamic information environments. Ultimately, an LLM output is only as good as its retrieved inputs (Shi et al.,, 2023). If pertinent information cannot be retrieved, no amount of LLM sophistication can compensate.

3 Capability Levels: A Five-Level Framework

While recent advancements in Retrieval-Augmented Generation (RAG) have delivered encouraging results, we argue that a more meaningful framing for enterprise use cases is: What kinds of questions can be reliably answered from private data? In enterprise settings, private data is not a free variable, it must be acquired, maintained, and protected. Much of the knowledge required for enterprise decision-making exists only in proprietary sources. Therefore, RAG capabilities should be measured not by end accuracy per se but by the depth and reliability of answers across increasingly complex data landscapes.

To standardize this perspective, we propose a five-level classification system organized around increasing complexity of reasoning and understanding required to support real-world enterprise tasks.

3.1 L1: Surface Knowledge of Unstructured Data

At L1, systems operate on unstructured text and perform basic retrieval using semantic similarity or keyword matching. They can answer factoid or lookup-style questions when the information exists explicitly in the source. However, they are unaware of any structural or contextual relationships among facts and cannot resolve ambiguity or synthesize across documents. These systems are effective for simple use cases such as semantic search, FAQ answering, and document summarization, where the required knowledge is shallow and localized.

3.2 L2: Surface Knowledge of Multifaceted Data

L2 systems can process a wider range of multi-modal and multi-structural data types, including tabular data, knowledge graphs, text, images, and metadata. They incorporate awareness of document structure and modality, enabling better matching for more complex formats. However, their responses remain surface-level: they retrieve based on local matches without deeper inference or understanding. These systems are context-aware, but not reasoning-capable. They are suitable for moderately structured content like contracts, handbooks, or technical manuals, where surface context is helpful but not sufficient for deep understanding.

3.3 L3: Implicative Knowledge of Multifaceted Data

L3 introduces reasoning capabilities. Systems at this level can synthesize information across multiple sources, infer relationships between disparate data points, and respond to queries that involve ambiguity, inconsistency, or incompleteness. They move beyond surface matching to link entities, draw conclusions across documents, and bridge questions with their many possible interpretations. This level is essential for enterprise scenarios involving diagnostic analysis, compliance validation, or decision support where answers depend on diverse, distributed, and contextualized knowledge.

3.4 L4: Reflective and Reasoned Knowledge

At L4, the system becomes self-aware in its reasoning process. It can reflect on the sufficiency and reliability of its own answers, identify missing context, and revise its retrieval or reasoning strategy accordingly. It may dynamically invoke tools (e.g., calculators, database queries, visual parsers) or select from multiple data sources depending on the task. This level introduces true agentic behavior: the ability to plan, adapt, and reason recursively. It supports high-stakes applications such as enterprise copilots, compliance agents, and automated report generation where accuracy, traceability, and adaptive reasoning are non-negotiable.

3.5 L5: General Intelligence

L5 represents an aspirational future state: a system capable of general-purpose reasoning across any domain, modality, or task type. It would possess the capacity to generalize knowledge, solve novel problems, and operate with minimal task-specific tuning or prompting. While it remains beyond current system capabilities, it serves as a direction for long-term research and system design.

Capability		Data Types Handled	Knowledge Capability Description	Example Tasks
L1	Surface Knowledge of Unstructured Data	Primarily text-based unstructured sources	Retrieves explicit facts from text using keyword or semantic similarity; no structural or multi-source reasoning.	FAQ answering, document search, basic summarization
L2	Surface Knowledge of Multifaceted Data	Unstructured + semi-structured (tables, metadata, images)	Retrieves from diverse formats with structural context awareness, but answers remain surface-level without inference or synthesis.	Table lookup, contract clause search, metadata-based filtering
L3	Implicative Knowledge of Multifaceted Data	Unstructured + semi-structured + structured (databases, graphs)	Synthesizes across sources and modalities, infers relationships, and handles ambiguity, incompleteness, and inconsistencies.	Root cause analysis, compliance checks, product comparisons
L4	Reflective and Reasoned Knowledge	Multi-modal (text, tables, APIs, images, code)	Adapts retrieval strategies, orchestrates tools, and verifies or revises answers for high-stakes tasks: self-aware reasoning	Regulatory report generation, enterprise copilots, design validation
L5	General Intelligence	Any domain, modality, or data structure	Aspirational: domain-agnostic, autonomous reasoning and general-purpose problem-solving across all modalities.	Open-ended research, autonomous decision-making, novel problem solving

Table 1: Five-Level Knowledge Capability Framework for Enterprise RAG Systems. Classification of systems based on the depth of reasoning and synthesis achievable from enterprise data, along with representative tasks at each level.

4 Methodology: Enabling L4 Systems

Achieving L4 (reflective and reasoned knowledge) requires a fundamental departure from traditional dense-vector search and static chunk-based retrieval. While L1 through L3 are sufficient for simple lookup and context-aware or even cross-source synthesis, enterprise environments often demand capabilities that are dynamic, multi-modal, and self-correcting. L4 systems must retrieve relevant information but also evaluate, adapt, and orchestrate their reasoning to ensure accuracy and completeness in mission-critical settings.

To meet these demands, an L4 system must:

• Accurately represent and unify complex structured, semi-structured, and unstructured data

• Retrieve context across multiple semantic, structural, and metadata dimensions

• Dynamically adapt retrieval and reasoning strategies to the query’s intent

• Seamlessly orchestrate multiple tools, models, and knowledge views

• Self-reflect, re-plan, and re-execute retrieval steps to “connect the dots”

We introduce Corvic AI, a platform designed from the ground up to operate at L4. Corvic AI integrates three pillars, (1) Structure-Aware Data Representation, (2) Mixture of Spaces, and (3) Adaptive Chain of Actions, transforming static pipelines into highly accurate agentic reasoning systems.

4.1 Structure-Aware Data Representation

Enterprise knowledge ecosystems rarely exist in a single format. They blend unstructured content (e.g., manuals, reports, narrative text) with structured elements (e.g., tables, forms, relational databases, graphs). L4 performance depends on a unified representation that captures the multi-faceted nature of the data.

The Structure-Aware Data Representation approach parses each document into an enriched intermediate form encoding:

• Document hierarchies, e.g., section, subsection, heading, paragraph hierarchies

• Embedded tables, lists, and field-level forms

• Local and global metadata, e.g., document type, version, author

• Cross-references, e.g., “See Table 1”, “as shown above”

This structure-aware representation bridges the gap between unstructured and structured modalities, enabling precise section or field-specific retrieval, cross-format linking, and context mapping across disparate data types. It is modular and extensible, and it forms the foundation for the techniques that follow (as shown in Fig. 2).

4.2 Mixture of Spaces

Traditional RAG systems (as shown in Fig. 1) compress all content into a single semantic vector space, losing critical structural and contextual cues. The Mixture of Spaces (MoS) approach builds multiple parallel representations of the same document:

• A semantic space for meaning-rich embeddings

• A structural space modeling layout, hierarchy, and relationships

• A metadata space encoding titles, tags, and annotations

This multi-view indexing allows the system to retrieve relevant context through different pathways. For example, if a semantic search misses a passage, a structural or metadata search may still retrieve it based on its heading or relational position. This redundancy increases both recall and precision. This is key for high-stakes decision-making.

4.3 Adaptive Chain of Actions

Most RAG pipelines follow a rigid retrieve, augment, and generate sequence. In contrast, the Adaptive Chain of Actions (ACoA) framework dynamically assembles query-specific retrieval and reasoning plans.

Adaptive Chain of Actions can:

• Select the optimal retrieval spaces based on query intent

• Sequence multiple retrieval and enrichment steps (e.g., schema lookup $\rightarrow$ graph traversal $\rightarrow$ synthesis)

• Integrate specialized tools for computation, visualization, or external data access

By adapting its strategy on the fly, ACoA supports reflective reasoning: the ability to detect gaps, revise plans based on discovered knowledge, and re-execute and refine steps until the answer meets completeness and reliability criteria.

In combination, Structure-Aware Data Representation, Mixture of Spaces, and Adaptive Chain of Actions enable Corvic AI to operate at L4 (reflective and reasoned knowledge), bridging structured and unstructured sources, adapting to varied enterprise queries, and delivering accurate, explainable, and high-trust outputs (as shown in Fig. 2). The following section evaluates Corvic AI’s performance against other leading RAG platforms.

5 Experimental Results

5.1 Datasets

Conventional benchmarks for Retrieval-Augmented Generation (RAG) typically evaluate performance on preprocessed, clean text chunks. This approach, however, is not representative of real-world enterprise scenarios, where information is predominantly stored in unstructured documents. For a more precise performance assessment on realistic enterprise tasks, our evaluation employs four datasets with source documents in PDF format. Each dataset is paired with a corresponding question corpus, formulated to probe for information embedded within the documents’ complex structure .

• DelucionQA: The DelucionQA dataset (Sadat et al.,, 2023) is a benchmark for question answering, using the Jeep 2023 Gladiator Car manual as its knowledge base. DelucionQA provides a challenging dataset designed to evaluate a system’s ability to answer specific, technical questions based on a complex, real-world document.

• FinTabNet: The FinTabNet dataset is a large-scale collection of PDFs from the financial earnings reports of Fortune 500 companies (Zhao et al.,, 2024; Zheng et al.,, 2020). Released as an open-source resource by IBM Research, it is freely available on the IBM Developer Data Asset Exchange.

• DMV Handbooks: The DMV Handbooks dataset is a Corvic AI-created dataset comprising the official DMV handbooks from all 50 US states. This collection of 50 long-form, structurally complex PDFs serves as the source for answering comprehensive sets of official state DMV test questions. It evaluates a system’s ability to perform precise information retrieval across a large and heterogeneous document collection, mirroring a common enterprise requirement.

• Architectural Manuals: The Architectural Manuals dataset is a Corvic AI-created dataset containing real-world complex architectural and engineering documentation. This includes CAD drawings, technical schematics, circuit diagrams, and building project blueprints that combine intricate 2D visual elements with embedded technical text. The dataset is designed to test a system’s ability to jointly process and reason over multi-modal inputs where precise visual–text alignment is critical.

Dataset	Modality	#Q	#Files	#Pages	Enterprise Use Case
DelucionQA	Text	184	1	384	Basic retrieval from manuals/product guides; customer support and FAQ automation.
FinTabNet	Text + Tables	1,000	591	591	Financial analysis from reports; risk assessment, KPI computation, and compliance reporting.
DMV Handbooks	Text + Images + Tables	135	50	4,689	Regulatory/compliance document search; onboarding, training, and policy lookup.
Architecture	Images + Diagrams + Embedded Text	43	8	36	Engineering/design review; reasoning over blueprints, CAD diagrams, and technical schematics.

Table 2: Datasets (questions, number of PDF files and pages) in increasing complexity. Complexity increases with modality variety, reasoning depth, and retrieval difficulty. The datasets are available for download at: https://huggingface.co/datasets/corvicai/delucionqa, https://huggingface.co/datasets/corvicai/FinTabNet_ComTQA, %****␣main.tex␣Line␣250␣****https://huggingface.co/datasets/corvicai/dmv-handbooks, and https://huggingface.co/datasets/corvicai/architectural-manuals.

	Dataset	Example Query	Ground Truth	Complexity
L1	DelucionQA	What is the DEF?	Diesel Exhaust Fluid	Simple question on simple text search; answer explicitly stated in plain text.
L2	FinTabNet	For PPL, which portfolio had the highest percentage of assets allocated to debt securities in 2015?	Growth Portfolio (Rationale: Debt Securities = 13%)	Direct question but requires searching and reading tables from unstructured data accurately to locate the answer.
L3	DMV Handbooks	What is the estimated BAC for a 120-pound woman in California after 2 drinks?	$\approx$0.11% (Rationale: BAC chart entry with 120 lb, 2 drinks)	Direct question but answer is not in plain text; requires combining table lookup with multimodal (chart) parsing and implicit knowledge generation.
L4	Architectural Manuals	Find an enclosure with three knockouts.	Enclosed Series KT7 Motor Controller (Rationale: From Figure 3, the enclosure with three circular knockouts)	Indirect question requiring deeper domain understanding; needs multimodal retrieval and reasoning across diagrams, labels, and text.

Table 3: Illustrative examples mapped to capability levels. The additional complexity column explains the reasoning complexity behind each example.

The examples in Table 3 illustrate the progressive complexity of retrieval-augmented generation (RAG) tasks as the capability level increases. At L1, such as the DelucionQA dataset, the question is straightforward and the answer exists verbatim in the text, requiring only basic semantic search over unstructured content. FinTabNet (L2) introduces structured elements, in this case, a table, which demand accurate parsing and retrieval of non-textual formats, but answers are still directly factual. DMV Handbooks (L3) raise the bar by requiring multi-modal comprehension: the system must interpret a chart, map its contents to the query, and apply implicit reasoning to extract the correct result. Finally, Architectural Manuals (L4) represent an indirect, domain-specific query where the answer is not explicitly stated in any single source; instead, it requires understanding technical terminology (“knockouts”), combining information from diagrams and text, and synthesizing a correct and contextually appropriate answer. This progression underscores how higher levels demand richer data representations, more sophisticated retrieval strategies, and reasoning capabilities that extend well beyond surface-level keyword or vector search.

5.2 Evaluation

For our experiments, we selected four representative implementations aligned with the capability levels. LangChain was used to implement an L1 (surface knowledge of unstructured data) system, focused on semantic search over unstructured text. OpenAI RAG and Azure AI Search represent L2 (surface knowledge of multifaceted data) systems, capable of handling unstructured and semi-structured data with surface-level structural awareness but without deep reasoning. Corvic AI is representative of an L4 (reflective and reasoned knowledge) system. It incorporates structure-aware representation, Mixture of Spaces, and Adaptive Chain of Actions for reflective and adaptive reasoning.

To ensure consistent comparison of response quality, we used gpt-4.1 for the final answer generation and summarization step. Depending on the system, there are a variety of parameters that affect retrieval. To obtain representative behavior, we used the default retrieval settings. LangChain and Corvic AI use OpenAI’s text-embedding-3-large embeddings, while OpenAI RAG uses its native embedding solution. LangChain uses ChromaDB as the vector database. For Azure AI Search, we used the default parameters which use OpenAI’s text-embedding-3-small model for embeddings.

For evaluation, we employed an LLM-as-a-judge (Zheng et al.,, 2023) approach using the ragas library (Es et al.,, 2024), leveraging gemini-2.0-flash as the evaluation model to score the outputs. This choice ensures an unbiased assessment, as the judge model is from a different family than the generation model used in the pipelines.

Note: In this study, our primary goal was to measure the performance advantage that L4 capabilities bring to a knowledge search and management use case primarily involving unstructured data; future work will extend this evaluation to multi-structured datasets along with unstructured ones.

Framework	DelucionQA	FinTabNet	DMV Handbooks	Architectural Manuals
LangChain	69.83	38.23	64.63	36.63
Azure AI Search	67.11	41.75	51.30	40.12
OpenAI RAG	67.53	52.55	71.85	58.72
Corvic AI	82.74	63.83	79.44	65.12

Table 4: A performance comparison of four AI systems across different datasets. Scores are reported as accuracy percentages (%) with the highest score in each row shown in bold.

Figure 3: DelucionQA: Manual-like Q&A over unstructured text and images. Corvic AI leads across basic retrieval questions on a single long manual.

Figure 4: DMV Handbooks: Mixed text, tables, and images. Corvic AI widens the gap as queries require cross-section references and light reasoning.

Figure 5: FinTabNet: Table-centric questions. Corvic AI outperforms on multi-step, numeric, and schema-aware reasoning over financial tables.

Figure 6: Architectural Manuals: Diagrams with embedded text and layout noise. Corvic AI shows gains on visual–text fusion and context-heavy queries.

Figure 7: Performance of systems aligned to capability levels across four datasets. Bars show comparative scores per system for each dataset (higher is better). Across all corpora, the L4 (reflective and reasoned knowledge) system, Corvic AI leads; L2 varies by implementation (e.g., retrieval strategy, embeddings, re-rankers) and gaps widen with dataset complexity (DelucionQA $<$ DMV $<$ FinTabNet $<$ Architecture).

5.3 Key Observations

Across all four datasets, Corvic AI, which is designed to solve L4 (reflective and reasoned knowledge) problems, delivers the strongest accuracy, even when the corpora are primarily unstructured. The combination of Structure-Aware Data Representation, Mixture of Spaces, and Adaptive Chain of Actions consistently recovers evidence that single-space pipelines miss (e.g., passages discoverable via headings or layout structure rather than pure semantics) and then verifies or augments that evidence through iterative retrieval. In practice, this manifests as higher groundedness and fewer omissions on long, layout-heavy PDFs, where relevant content is often contextually implied by section hierarchy or table locality rather than surface phrasing.

Within the same capability tier, performance can vary markedly, e.g., among the two L2 (surface knowledge of multifaceted data) systems (Azure AI Search vs. OpenAI RAG). This divergence highlights the impact of design choices: vector vs. keyword vs. hybrid retrieval (and their weighting), domain vs. general-purpose embedding models, cross-encoder vs. bi-encoder re-rankers, and the summarization LLM’s length control and citation style. Small configuration shifts can change which passages are retrieved, how they are ordered, and what ultimately makes it into the context window. This, in turn, leads to materially different answer quality without changing the nominal “level” of the system.

Dataset complexity acts as an amplifier. On simpler, manual-like material (DelucionQA), the gap between L1 (surface knowledge of unstructured data) and higher level systems is narrower because many answers are explicitly stated and semantically proximate. As structure and reasoning demands grow, from the simpler DMV (mixed modalities and long-range references) dataset to the increasingly more complicated FinTabNet (table-centric, numeric, and multi-step computation) and Architectural Manuals (visual–text fusion with layout noise) datasets, the advantage of Corvic AI widens compared to L2 systems on average. Multi-space retrieval and reflective planning help recover table cells, respect schema and column context, and follow cross-figure references, while systems designed for lower levels suffer recall drops or over-summarization errors.

A related pattern is robustness. Corvic AI not only achieves higher accuracy but also exhibits smaller swings across datasets. Reflective re-planning reduces brittleness to corpus idiosyncrasies (e.g., varied headings, inconsistent table markup, OCR artifacts) by retrying with alternate spaces or tools when evidence is sparse. In contrast, systems designed for L1 and L2 show larger sensitivity to modality mix and layout density because they rely more heavily on a single retrieval view and fixed ordering.

Finally, the error taxonomy differs by level. LangChain (L1) tends to miss cross-section or cross-document cues and may retrieve semantically similar but contextually wrong passages. L2 systems improve recall but are brittle: changes in hybrid weights or re-ranker choice can flip top passages, and table/figure evidence is often underutilized. Corvic AI mitigates both classes of errors by (1) retrieving through multiple views (semantic, structural, metadata), (2) verifying coverage against the question intent, and (3) iterating when gaps are detected, which collectively reduces hallucinations and incomplete answers on enterprise-style documents.

6 Discussion

Our results show that techniques for addressing L4 (reflective and reasoned knowledge) problems deliver a clear and consistent advantage for enterprise knowledge search and management even when corpora are predominantly unstructured. The architectural ingredients behind L4 (Structure-Aware Data Representation, Mixture of Spaces, and Adaptive Chain of Actions) translate into higher accuracy and more reliable grounding across diverse datasets, while lower levels exhibit wider swings.

The study also clarifies how to interpret “levels” versus “implementations”. A level specifies a capability envelope. Where a concrete system lands within that envelope depends on design choices, e.g., retrieval strategy, embeddings, re-rankers, summarization LLM. Thus, the classification provides the ceiling and direction of travel, and implementation determines realized performance within a level.

Finally, as task and data complexity rise, the value of higher levels grows. This suggests a pragmatic roadmap: organizations starting with simpler use cases can begin at lower levels but should plan upgrades toward L4 as modality mix, reasoning depth, and reliability requirements increase.

Scope. This study focused on question answering on datasets primarily comprised of unstructured-data. Future work will extend to structured and multi-modal corpora and include L3 baselines and targeted ablations to map accuracy–latency–cost trade-offs more completely.

7 Conclusion

We introduced a five-level RAG classification that organizes system capabilities from L1 (surface knowledge of unstructured data) through to L4 (reflective and reasoned knowledge) and the aspirational L5 (general intelligence). This is intended both as a methodological framework and as a practical guide for aligning enterprise requirements with implementation choices.

At lower levels, results reveal that choices of retrieval strategy, embeddings, re-rankers, and summarization LLMs can shift outcomes. This underscores the value of system classification for setting appropriate expectations.

Corvic AI consistently outperforms representative L1 and L2 systems on knowledge search and management tasks, even on predominantly unstructured datasets. Since Corvic AI was designed to solve L4 problems, this highlights the leverage of addressing complex problems upfront.

As enterprises adopt RAG for mission-critical workflows, framing the problem correctly becomes essential for determining the right system capabilities—and Corvic AI exemplifies how higher level problem solving can deliver measurable enterprise value today.