NKK argues that much of RAG's headroom sits upstream of the model; no public source sizes the share. Four principal NKK models for readiness, pattern choice, trust and failure triage, built on cited evidence.
Written for: CIOs, chief data officers, heads of AI, platform and knowledge-management leads
12 chapters, 65 cited sources, 20 models and charts. The first 2 chapters are open; the rest unlock with the form.
accuracy of straightforward RAG on CRAG; most advanced LLMs reached 34% or less without retrieval
hallucination rate of leading RAG-based legal research tools in the 2024 preprint (the 2025 journal version says more than 17%)
top-20 retrieval failures after contextual embeddings, contextual BM25 and reranking, down from 5.7% (Anthropic's own data)
Knowledge lives in drives, inboxes and chat. Nobody owns it, and parsing is whatever the tool's default does.
Each corpus has a named owner and a place in a register. Parsing has been checked against the original documents on a sample.
Classification and access rights travel with every chunk and are enforced when the query runs. Each source has a freshness rule. A first golden set, including questions the corpus cannot answer, has been run before launch.
Hybrid retrieval with a keyword baseline, reranking where it pays, and a golden set grown from real queries and gated at every release.
Every answer traces to a source version and owner. Drift and staleness are monitored, and graph or agentic retrieval is used only where measurement shows it pays.
In NKK's view, much of the headroom in production RAG sits upstream of the model, in two different places: retrieval engineering, and the governance of the content behind it. No public source measures what share of RAG failures each causes. The public figures show that RAG alone does not guarantee accuracy; they do not locate the cause. On Meta's CRAG benchmark, straightforward RAG improved accuracy only to 44%, while most advanced LLMs reached 34% or less without retrieval [1]. A 2024 Stanford preprint found that leading commercial legal research tools built on RAG hallucinated between 17% and 33% of the time; the peer-reviewed 2025 version words the headline as "more than 17%" [2]. Of these three figures, only the Anthropic result isolates an upstream change: in Anthropic's own tests, context-enriched chunks, a keyword index and a reranker cut top-20 retrieval failures from 5.7% to 1.9% [3]. That is vendor-reported retrieval engineering, not evidence on ownership, permissions or freshness.
This paper argues that knowledge readiness is a leading constraint on enterprise RAG, and that the plumbing around it is commoditising fast. It offers four principal NKK models, the Knowledge Readiness Ladder, the Retrieval Pattern Selector, the Retrieval Trust Chain and the Failure-to-Control Map, plus supporting tools such as a scorecard and an evaluation loop. All of it is NKK opinion, not validated prediction, and NKK has no client results to report. Public evidence is thin: chunking evidence comes from vendor reports and a single peer-reviewed study, GraphRAG gains are shown on a handful of corpora, and several permission features are still in preview.
accuracy of straightforward RAG on CRAG; most advanced LLMs reached 34% or less without retrieval
hallucination rate of leading RAG-based legal research tools in the 2024 preprint (the 2025 journal version says more than 17%)
top-20 retrieval failures after contextual embeddings, contextual BM25 and reranking, down from 5.7% (Anthropic's own data)
A common reflex when a team misses its quality bar is to reach for the same levers: a larger model, a longer prompt, a new embedding model. These are the visible parts of the system. The invisible part is the supply chain that fills the index: which documents exist, how they were converted to text, who may see them, how old they are, and whether anyone can tell when an answer is wrong.
The most useful public map of where RAG systems break is an experience report from Deakin University. Drawing on three case studies (a research tool, an AI tutor and biomedical question answering), the authors identified seven failure points [4]:
In NKK's reading, which is our interpretation and not the paper's, the first three are decided before generation starts, by what the corpus contains, how it is ranked and how it is assembled. Only FP4 to FP7 are mainly generation behaviour. That matters for budgeting: a team that spends its effort on the generator is working on four of the seven failure points and leaving the first three untouched.
The first failure point deserves a second look because it is easily mistaken for a model flaw. A system that answers a question the corpus cannot support is not hallucinating in the sense of failing to read; it is failing to know what it does not know. Google researchers studying "sufficient context" found that larger models often output an incorrect answer instead of abstaining when the retrieved context is insufficient, while smaller models hallucinate or abstain often even when the context is sufficient [5]. They proposed selective generation using a sufficiency signal and reported that it improved the fraction of correct answers among the responses given by 2% to 10% for Gemini, GPT and Gemma models [5]. Separately, the RGB benchmark found that models struggle with negative rejection (declining when the answer is not in the documents) and with integrating information across documents [6]. The practical conclusion is that abstention is a property of the whole system and has to be tested with questions the corpus cannot answer.
Three gaps matter for anyone building a business case. First, no primary source we could open gives a reliable share of enterprise RAG pilots that fail or reach production, or the share of RAG failures caused by knowledge supply as against the model. That is why the thesis of this paper is framed as NKK's view. Widely repeated analyst statistics on "AI-ready data" could not be traced to a primary page, so this paper does not use them. Second, the Stanford AI Index reports that global corporate AI investment reached $581.7 billion in 2025, up 130% [7], but we found no breakdown of how much goes to knowledge foundations, and we would treat anyone who claims a precise figure with caution. Third, the strongest controlled evidence is on public benchmarks, and the authors of the newest enterprise-style benchmark warn that benchmark performance does not necessarily transfer to enterprise retrieval [8].
Every figure and regulatory statement in this piece traces to a numbered source below. Links open the original publisher in a new tab.
A bigger window did not solve the problem. Why context is a budget, what the evidence shows and does not, and NKK's layer stack and triage matrix, with a ledger and scorecard to apply them.
Debug RAG bottom-up: corpus, candidate recall, rank and cut, context use, answer shape, authorization. A sourced guide to chunking, hybrid search, reranking, evaluation and permissions, with five NKK models.
A vendor-neutral reference architecture for enterprise agents: NKK's nine-layer stack, a pattern-selection matrix, a blast-radius gate, and the evidence on reliability, state, control and cost.
