Company knowledge needs more than a vector database

Private RAG that respects documents, permissions and uncertainty.

Retrieval-augmented generation can ground an AI response in approved company sources. The hard parts are document quality, permissions, retrieval evaluation and human review - not simply installing a chat interface.

Decision in one minute

Use RAG when answers should be grounded in a controlled source set. Do not use it as a substitute for source governance or professional judgement.

Three-quarter supplier render of a 4U OEM multi-GPU rack server
Three-quarter supplier render of a 4U OEM multi-GPU rack server
OEM platform reference render. It is not evidence of a completed customer build or final specification. OEM supplier reference image. Written reuse permission pending.
OEM platform reference render. It is not evidence of a completed customer build or final specification.

Buyer field note

Let the work and the room reject the wrong machine.

01 / Fit
Start with the accepted task, not a chassis or GPU headline.
02 / Site
Check power, cooling, noise, network and maintenance access.
03 / Alternative
Keep cloud, colocation or a smaller system in the decision.
Diagram showing an approved request, a local service, an approved store and a policy-controlled data path
Diagram showing an approved request, a local service, an approved store and a policy-controlled data path
A private deployment starts with the permitted data path, access policy and logging boundary. Original explanatory plate. It sets out a decision method, not a measured result.
A private deployment starts with the permitted data path, access policy and logging boundary. Original explanatory plate. It sets out a decision method, not a measured result.

A transparent data path

Documents are parsed, split, embedded and stored for retrieval. At question time, relevant passages are passed to the model and the answer should cite them.

Each stage needs a named owner, retention rule and permission model. A local deployment can keep that path close, but it does not correct poor source data.

Capacity budget

Memory demand has more than one layer

Model
Weights and quantisation.
Context
Working input and cache.
Concurrency
Simultaneous active work.
Headroom
Operations and change allowance.
The bands are a checklist, not a scale. Exact memory belongs to a named model and test. Decision aid for: A transparent data path

Connect the category to the physical and operating system.

Platform views and technical plates show what must be tested beyond a product label or aggregate specification.

Diagram of an evidence pack containing an asset schedule, burn-in record, health readings, workload test and admin guide
Diagram of an evidence pack containing an asset schedule, burn-in record, health readings, workload test and admin guide
A credible handover records the supplied assets, checks, operating evidence, instructions and unresolved items. Original explanatory plate. It sets out a decision method, not a measured result.
A credible handover records the supplied assets, checks, operating evidence, instructions and unresolved items. Original explanatory plate. It sets out a decision method, not a measured result.
GPU server remote management dashboard with system status, access logs and sensor monitoring panels
GPU server remote management dashboard with system status, access logs and sensor monitoring panels
Supplier screenshot of the platform management interface. The final management features and access policy depend on the ordered system. OEM supplier reference image. Written reuse permission pending.
Supplier screenshot of the platform management interface. The final management features and access policy depend on the ordered system. OEM supplier reference image. Written reuse permission pending.

Permissions must survive ingestion

A single company-wide index can leak information between teams. The design must preserve access groups or use separate collections where the source systems require it.

The user interface should show sources and make it obvious when retrieval confidence is weak.

Site boundary

The room is part of the specification

  1. Electrical Circuit, protection and UPS policy.
  2. Thermal Airflow, room heat and cooling route.
  3. Network Access, bandwidth and isolation.
  4. Service Rack space, noise and maintenance access.
A technically suitable server is still the wrong purchase when the operating site cannot support it. Decision aid for: Permissions must survive ingestion

Evaluate retrieval and answer quality separately

The system can retrieve the wrong passage even when the language model writes a fluent answer. Conversely, the right passage can be misinterpreted.

A representative question set should score retrieval, citation, answer support and refusal behaviour before wider access.

Procurement record

Make the route to acceptance inspectable

  1. Fit brief Workload and alternatives recorded.
  2. Reference test Conditions and results retained.
  3. Exact build Bill of materials and substitutions visible.
  4. Handover Owners, versions and exclusions signed off.
The evidence trail should explain why this platform was chosen and what remains outside the order. Decision aid for: Evaluate retrieval and answer quality separately

Hardware depends on usage, not document count alone

Ingestion can be a batch task. Day-to-day GPU demand is driven more by model, context, concurrent questions and response target.

Storage size, backup and document update frequency still matter, but millions of file names are not a useful server-sizing headline by themselves.

Questions answered

Straight answers to common questions

What is a private RAG server?

A server that runs retrieval, vector storage and a language model inside a customer-controlled environment so staff can ask questions over approved documents.

Will RAG stop hallucinations?

No. It can ground answers and expose sources, but retrieval and generation can still fail. Evaluation, citations, refusal behaviour and human review remain necessary.

Can it connect to SharePoint or Google Drive?

Connectors can be scoped separately. The initial appliance baseline does not include every integration or permission mapping by default.

Put the claim to work

Turn this guidance into a testable requirement.

The brief asks about workload and operating conditions - not just budget.