Extract, and measure coverage.
Text from every format, with optical character recognition where needed, and a recorded share of documents that yielded usable text.
The demo worked.
The corpus
did not.
Enterprise retrieval fails on the unglamorous parts: scanned documents, six source systems, permissions that change hourly, and a corpus that grows daily.
READ WHAT CHANGES ↓SOURCES · FRESHNESS · SCALE
01 / DEFINITION
The short answer first.What is enterprise RAG?
Enterprise RAG is retrieval-augmented generation deployed across an organisation's real document estate: multiple source systems, inconsistent formats, inherited permissions, continuous change and volumes large enough that ingestion and freshness become engineering problems in their own right.
The gap between a prototype and an enterprise deployment is rarely the model or the vector store. It is the ingestion pipeline, the permission model and the operational commitment to keeping an index current — none of which are visible in a demonstration built on a clean folder of PDFs.
02 / WHAT ACTUALLY BREAKS
The list nobody plans for.| Problem | Why it appears | Consequence if ignored |
|---|---|---|
| Scanned documents | A real estate contains images of text, not text | Silently missing content; answers with no sources |
| Format sprawl | Email, spreadsheets, presentations, archives, legacy formats | Partial extraction that nobody notices |
| Tables and structure | Naive extraction destroys tabular meaning | Confidently wrong numerical answers |
| Duplicates and versions | The same document exists in five places at three revisions | The model cites a superseded version |
| Permission inheritance | Rights live in the source systems, not the index | The index becomes the widest access path in the organisation |
| Freshness | The corpus changes continuously | Answers drawn from material that has been superseded or deleted |
Scanned material is the most consequential because its failure is silent. A pipeline that cannot read an image-based PDF produces no error; the document is simply absent from every answer, and the gap is invisible until someone notices a citation that should have existed. Measuring extraction coverage — what share of the corpus yielded usable text — is a basic ingestion metric and is rarely collected.
Version confusion is the second. In document estates where superseded drafts sit alongside executed versions, retrieving the wrong one produces an answer that is fluent, sourced and wrong. Version and status metadata belong in the index, and the retriever should prefer current material explicitly rather than by accident of ranking.
03 / INGESTION AS A PRODUCT
The component that determines the ceiling.Normalise and deduplicate.
One canonical record per document, with versions linked rather than duplicated as independent entries.
Chunk with structure in mind.
Boundaries that respect headings, clauses and tables, with overlap. This sets the recall ceiling for the whole system.
Carry permissions and provenance.
Access metadata, source, version, status and date on every chunk, from the first version of the pipeline.
Step four is the one that cannot be retrofitted cheaply. An index built without access metadata has to be rebuilt to become identity-aware, and by the time that is discovered the pipeline is usually processing millions of chunks. Carrying permissions from day one costs a schema field. See RAG access control.
Ingestion is also where most of the ongoing operational load sits. New documents, changed documents, deleted documents, re-processing after a chunking improvement, and re-embedding after a model upgrade are all continuous activities, and a full re-index of a large corpus is a scheduled project rather than a command.
04 / WHAT TO BUILD FIRST
Sequencing an enterprise deployment.One source system, one user group.
A narrow, permissioned slice that reaches real users, rather than a complete index that reaches nobody.
The evaluation set before the second source.
Without it there is no way to tell whether adding a source improved anything or diluted retrieval.
Coverage and freshness dashboards.
Share of the corpus successfully extracted, indexing lag, empty-retrieval rate. These predict user-visible quality better than any model metric.
Adding sources is not monotonically beneficial. Each new source adds candidate passages that compete for a fixed context budget, and a low-quality or redundant source can reduce answer quality measurably. That is precisely the kind of change an evaluation set makes visible and intuition does not — which is why the measurement belongs before the expansion. See RAG evaluation and private AI deployment.
05 / QUESTIONS
Asked when retrieval moves beyond a pilot.Why does an enterprise RAG prototype degrade in production?
+
Usually ingestion and volume. A prototype runs on clean, curated documents; production includes scans, legacy formats, duplicates, superseded versions and material the pipeline silently fails to read. Extraction coverage is the first number to measure.
How do you handle scanned documents?
+
Optical character recognition in the ingestion pipeline, with a recorded quality signal, because poor recognition produces plausible nonsense that is worse than an omission. Track the share of the corpus that yielded usable text and treat a low figure as a defect.
Should everything go into one index?
+
Not automatically. More sources mean more competition for a fixed context budget, and a redundant source can reduce quality. Partitioning also makes permission isolation structural rather than filter-dependent. Measure the effect of each addition.
How current does an index need to be?
+
As current as the decisions taken from it require. For a matter a person is actively working on, minutes; for a historical archive, days. Choose per source, and make the indexing lag a monitored metric rather than an assumption.
What is the realistic operational load of enterprise RAG?
+
Continuous: ingestion for new and changed documents, deletion propagation, periodic re-chunking and re-embedding as the pipeline improves, plus permission synchronisation. It is an operated service, not a project that completes.
Make the corpus
work.
Privilege AI builds retrieval and knowledge systems, including the ingestion and permission work that decides whether they hold up.