People search across too many places
Policies, manuals, case records and product information may sit in shared drives, business systems and separate databases.
Enterprise RAG & AI Knowledge Systems
Turn company knowledge into answers teams can check
Connect approved documents, databases and internal systems to AI assistants that retrieve relevant evidence, cite their sources and respect existing access permissions.
RAG can improve source grounding, but it does not guarantee that every answer is correct. We define citation, refusal, evaluation and human-review controls according to the workflow.
Illustrative evidence flow
Permission-aware retrieval before generation
Only authorised evidence should enter retrieval and the model context for this request.
- 01
Signed-in user
The request carries an authenticated identity and role.
- 02
Permission check
Source and tenant filters are applied before retrieval.
Approved sources includedRestricted sources excluded - 03
Hybrid retrieval
Semantic and lexical search find and refine relevant passages.
- 04
Cited response
The assistant answers from retrieved evidence and shows its sources.
Source-backed answerThe current procedure requires approval before release. [1] Exceptions follow the escalation path. [2]
[1] Procedure §4.2[2] Escalation guide
Controls across the full flow
- Evaluation set
- Audit record
- Source freshness
Response policy
Answer · Decline · Escalate
The hard part is not connecting a model
Company knowledge is usually spread across files and systems with different owners, update cycles and access rules. A useful knowledge assistant has to work with those operational realities.
Sources change without a clear refresh path
An answer can look convincing while relying on an old file unless deletion, replacement and re-indexing are part of the design.
Access rules differ by user and document
A general search index is not enough when teams, customers or tenants are allowed to see different information.
There is no agreed test for a good answer
Without representative questions and expected evidence, teams cannot tell whether retrieval improved or simply changed.
Where a governed knowledge system helps
Start with a workflow where the source material, users, permissions and review owner can be defined. These are common starting points, not pre-built products.
Internal policy and procedure search
Help staff find the current instruction and open the cited source before acting.
Customer service and technical support
Retrieve approved troubleshooting, service and product information while retaining escalation paths.
Product and approved sales knowledge
Answer detailed questions from maintained catalogues, specifications and approved commercial material.
Onboarding and operational guidance
Guide staff through role-relevant procedures without exposing unrelated internal content.
Governed record lookup
Bring together permitted documents, database records and API results for a specific operational question.
Define what a reliable answer means before launch
Evaluation covers the evidence selected as well as the final wording. We agree the test questions, failure policy and review owner before a pilot is treated as production-ready.
- 01
Source and citation relevance
Check whether retrieved passages support the answer and whether citations point to the right section.
- 02
Representative question sets
Test routine, ambiguous, outdated and adversarial questions drawn from the intended workflow.
- 03
Insufficient-evidence behaviour
Require the assistant to decline or escalate when approved sources do not support an answer.
- 04
Permission and tenant tests
Confirm that users cannot retrieve or infer content outside their role, organisation or record scope.
- 05
Source lifecycle checks
Test updates, deletions, replacements and re-indexing so superseded content is removed from use.
- 06
Operating measures
Monitor retrieval quality, response time, model usage and cost within the agreed service boundary.
Illustrative access scenario
Permissions are enforced before content reaches the model
Identity, source ownership and query-time filters work together. The model should receive only the evidence this user is permitted to retrieve for this request.
- OAuth2, SAML or Active Directory integration
- Query-time metadata filtering
- Tenant and document-level separation
- Logged evidence and response decisions
User context
Operations manager · authenticated staff role
Available to this role
Current operating procedures and approved product manuals
Excluded from retrieval
HR files and customer records outside the assigned scope
Permitted result
An answer based on allowed passages, with citations and an audit record
Architecture detail
A retrieval stack selected around your data and controls
The exact components depend on document quality, update frequency, language, scale, latency and deployment requirements. The following patterns can be combined where they fit the use case.
01Document and data ingestion
Automated ETL can parse PDF, DOCX, JSON and API payloads, apply layout-aware OCR, extract metadata and use recursive or sliding-window chunking. Each source keeps ownership, version and permission metadata.
02Hybrid retrieval
Dense vector retrieval can use HNSW or Flat indexing with cosine or inner-product distance. BM25 adds lexical matching, with Reciprocal Rank Fusion combining the candidate lists.
03Cross-encoder reranking
A secondary model, such as Cohere Rerank or BGE-Reranker, can reorder the top-k candidates before prompt assembly. This can reduce irrelevant context and unnecessary token use.
04Permission-aware access control
Enterprise identity can be connected through OAuth2, SAML or Active Directory. Query-time metadata filters apply tenant, role and document rules before retrieved text is added to the model context.
05Evaluation and operations
Versioned test sets, retrieval traces, citation checks, audit logs, source refresh jobs and usage monitoring make changes visible and reviewable after release.
Choose the model after the workflow and data boundary are clear
OpenAI, Claude, DeepSeek and Qwen can be assessed as generation or reasoning options. The decision is based on the selected deployment, provider terms and measured performance on the agreed question set.
OpenAI
Tooling, model capability and enterprise controls
Claude
Long-context and document-heavy workflows
DeepSeek
Cost and deployment options to assess
Qwen
Chinese-language and deployment options to assess
Selection criteria
- Data handling
- Language quality
- Latency and cost
- Context and tool support
- Hosting constraints
Move from a bounded pilot to an operated system
A pilot should test one useful workflow with real permissions and representative questions. Broader access comes after the evidence, failure handling and operating owner are understood.
- 01
Assess sources and permissions
Map users, questions, systems, data owners, sensitivity and update paths.
- 02
Build a limited pilot
Ingest an approved source set and connect one practical user journey.
- 03
Evaluate retrieval and answers
Measure evidence relevance, citations, refusals, access rules and response quality.
- 04
Integrate with the workflow
Connect identity, applications, APIs, review queues, logs and support processes.
- 05
Monitor and improve
Review source changes, failed questions, usage, latency and cost before expanding scope.
What the system still needs people to manage
RAG changes how evidence is found and presented. It does not remove responsibility for source quality, permissions or decisions made from an answer.
- A cited response can still misread a source or omit relevant context. High-impact decisions need an appropriate human review step.
- If approved sources are incomplete, contradictory or out of date, the response may also be incomplete, contradictory or outdated.
- Access control depends on correct identity, metadata and source permissions. These controls need testing whenever roles or data connections change.
- Model and retrieval behaviour can change with configuration, source content or provider updates, so evaluation continues after launch.
Private AI, cloud or hybrid deployment
Choose the data boundary before choosing a model. Private hosting alone does not establish security: permissions, logging and the data sent to every external service still need review.
| Approach | Data and model boundary | Operating responsibility |
|---|---|---|
| Hosted cloud model | Approved retrieval context is sent to a model provider. Check retention terms, processing location and permitted data. | Provider operates the model; your project still needs connector, permission, usage and support controls. |
| Private deployment | Run retrieval and the selected model within an agreed private environment. Verify that telemetry and supporting services respect that boundary. | Plan compute capacity, model updates, backups and evaluation. Dedicated infrastructure can add cost even when usage is low. |
| Hybrid deployment | Keep selected sources or retrieval services private and send only approved context to a hosted model. Confirm exactly what leaves the environment. | Maintain both sides of the connection, including filtering, identity, network failures and provider changes. |
FAQ
Frequently asked questions
What affects the cost of an enterprise RAG system?
Document parsing and cleanup, source connectors, permission mapping and evaluation shape the initial scope. Running costs include indexing updates, model usage, storage, monitoring and support. Test representative questions and document formats before sizing a wider deployment.
What is retrieval-augmented generation?
Retrieval-augmented generation, or RAG, retrieves relevant information from approved sources and supplies it to an AI model as context for an answer. It can make the answer easier to trace, but it does not guarantee correctness.
Which sources can an enterprise knowledge system use?
The agreed scope can include PDF and DOCX files, structured data, databases, APIs, and existing knowledge repositories. Each source needs suitable parsing, metadata, update, deletion, and access-control handling.
Can the system respect existing user permissions?
Yes, when identity, source permissions, metadata, and index boundaries are mapped and tested correctly. The exact approach depends on the organisation's identity provider, repositories, tenant model, and deployment environment.
Does RAG prevent incorrect AI answers?
No. Source quality, parsing, retrieval settings, permission data, and model behaviour can all affect an answer. We define citation, refusal, evaluation, logging, and human-review controls around agreed scenarios.
Which AI models can be connected?
A system can be designed for providers such as OpenAI, Claude, DeepSeek, or Qwen. Model selection and availability are confirmed against language, data handling, deployment, latency, cost, and integration requirements.
Talk through scope, timeline, and next steps
Start with the questions your team needs to answer
Bring a small set of representative questions, the source systems involved and the users who should have access. We can clarify the data boundary, pilot scope and evidence needed to judge the result.
The first step is a practical scope discussion, not a promise that AI can replace the workflow owner.
