There is a counterintuitive failure mode that AI teams hit when they try to improve their RAG pipelines: they add more data and accuracy gets worse. More documents, more context, more noise. The retrieval system surfaces less relevant content, the LLM has more to sift through and the responses degrade. 

Organizations that chase volume are solving the wrong problem.

AI accuracy is not a function of how much data you have. It is a function of how good that data is: how clean, how relevant, how well-prepared for retrieval and generation. The organizations pulling ahead on AI performance are not the ones with the largest document repositories. They are the ones with the most rigorously curated ones.

This distinction points to a capability that most AI programs have underinvested in: unstructured data curation, which is the process of discovering, scanning, filtering, enriching and governing unstructured content so that only high-quality, relevant and appropriately classified material reaches the AI pipeline. 

It’s the missing link between a mediocre AI deployment and a genuinely accurate one.

The accuracy paradox explained

When teams first build RAG systems, the logic seems obvious: more context equals better answers. So they ingest everything: SharePoint libraries, legacy file servers, S3 buckets, email archives, scanned documents. The retrieval corpus grows to millions of files and the accuracy numbers plateau or drop.

The reason is straightforward. Retrieval systems return the most semantically similar content, not the most correct content. If the corpus contains three versions of the same policy document — an outdated draft, a superseded version and the current one — retrieval may surface any of the three depending on the query. If it surfaces the wrong one, the generated answer is wrong, regardless of how good the underlying model is.

Garbage in, hallucination out. Not because the LLM is broken but because the data feeding it is.

What curation actually means

Curation is not editorial review. At enterprise scale, it cannot be. No team can read millions of documents and decide which ones belong in an AI corpus. Curation at scale means automated processes that systematically assess, filter and enrich content before it reaches the AI pipeline.

In practice, curation covers five distinct operations. Sensitivity scanning identifies which documents contain PII, regulated information, trade secrets or legally privileged content — content that should not be fed into an AI system without appropriate controls. Quality filtering removes content that is too thin, too old, too duplicative or structurally malformed to contribute meaningful signal. Deduplication eliminates near-duplicate documents that inflate the corpus without adding information. Metadata enrichment attaches structured context — document type, business domain, owner, data classification, creation date, version status — that allows retrieval systems to make smarter decisions about relevance. Relevance scoring assesses, at a document or chunk level, how likely a piece of content is to contribute useful signals for the use cases the AI is supporting.

These are not optional steps. They are the difference between a retrieval corpus that works and one that undermines the AI system it is supposed to support.

The accuracy lift: What better curation delivers

Organizations that have systematically curated their retrieval corpus before deploying AI report meaningful improvements in response accuracy. 

The pattern is consistent: starting accuracy rates in the 75-85% range improve to the 90-95% range when the corpus has been properly prepared. It’s the difference between an AI tool that users trust and one they abandon.

The mechanism is not mysterious. When outdated documents are removed, retrieval surfaces current information. When duplicates are collapsed, the retrieved context is more coherent. When metadata is enriched, the retrieval system can filter by recency, document type or classification, returning not just semantically similar content but contextually appropriate content. When sensitivity is flagged, the pipeline can apply appropriate access controls rather than inadvertently surfacing regulated content to unauthorized users.

Better curation does not require a better model. It requires better data going into the model. The model does not change; the input quality does.

The discovery–curation–delivery workflow

Improving AI accuracy through unstructured data curation follows a sequence that is repeatable and scalable. Here is how it works in practice.

Step 1: Discover. The first step is knowing what you have. Most organizations do not. Unstructured data lives across SharePoint, S3 buckets, network file shares, legacy document management systems, email archives and a long tail of departmental storage. Before you can curate, you need a complete inventory: what exists, where it lives, who created it and when. Automated discovery crawls these sources and builds a content inventory without requiring manual cataloging.

Step 2: Scan. Once you know what exists, you need to understand its sensitivity and risk profile. Automated sensitivity scanning classifies documents by the type of content they contain: PII, health information, financial data, attorney-client privileged material, trade secrets. This classification determines what can flow into the AI pipeline without restriction, what requires access controls and what should be excluded entirely. Skipping this step is a compliance risk as well as a governance failure.

Step 3: Filter. With a complete, classified inventory, you apply filters that remove content that does not belong in the AI corpus. Outdated documents beyond their useful life. Superseded policy versions. Duplicate and near-duplicate files that add noise without adding signal. Low-quality content, including scanned documents that failed OCR, files too short to contain meaningful context, content in unsupported formats. The corpus that remains is smaller and materially better.

Step 4: Enrich. Filtering removes the bad. Enrichment improves the good. Metadata enrichment attaches structured attributes — document type, business domain, data classification, version status, owner team — that make the content more retrievable and more governable. Semantic tagging connects documents to the business concepts and use cases they are relevant to. Enriched content surfaces better in retrieval and gives the AI system more context for generating accurate responses.

Step 5: Deliver. The curated, enriched, governed content is delivered to the AI pipeline — the vector store, the retrieval system, the knowledge base — with appropriate access controls applied and metadata preserved. The governance layer remains active: when source documents change, the pipeline is updated. When new content is added, it goes through the same curation workflow before reaching the AI system.

Why automation is non-negotiable

An organization with 2 million documents in its content estate cannot curate them manually. The math does not work at any staffing level. Even at three minutes per document, manual review of 2 million files would require roughly 100,000 hours of work. That is not a process; it is a fantasy.

Automation is not an enhancement to the curation workflow. It is the prerequisite for having a curation workflow at all. Every step in the discovery–curation–delivery sequence must be automated to be viable at enterprise scale: automated discovery, automated classification, automated filtering based on configurable rules, automated metadata enrichment using NLP and automated delivery with change detection to keep the corpus current.

This is also why curation cannot be a one-time project. Data estates are not static. New documents are created daily. Existing documents are updated. Policies change and classifications must be updated accordingly. Curation must be a continuous process — a pipeline, not a project.

Collibra and Deasy Labs: The solution in practice

Collibra and Deasy Labs have built a joint solution that addresses the full discovery–curation–delivery workflow for unstructured data. Deasy Labs brings specialized capabilities for automated metadata extraction, sensitivity classification and semantic enrichment at scale. Collibra provides the data governance layer — the catalog, the lineage, the access controls and the policy framework — that ensures curated content remains governed and auditable throughout its lifecycle in the AI pipeline.

The combination addresses the core challenge: not just getting unstructured data into an AI system, but getting the right unstructured data into the right AI system with the right controls applied. The result is a retrieval corpus that is accurate, compliant and continuously maintained. One that improves AI accuracy rather than undermining it.

Learn more about how Collibra and Deasy Labs are unlocking the value of unstructured data for the AI era and what that means for organizations serious about AI accuracy and governance.

Collibra helps AI teams transform unstructured data for AI — moving from raw content ingestion to curated, governed knowledge pipelines that produce AI systems worth trusting. Discover how the Collibra and Deasy Labs solution can close the accuracy gap in your AI deployment.

Want to see Deasy Labs in action? Request a demo.