AI
AI Agents & Enterprise Knowledge
SERVICE ENGINEERING GUIDE
Give teams an assistant that can find approved knowledge and complete bounded tasks. We design the knowledge pipeline, permissions, tool access and human handover together, so an agent has a clear operating boundary.
What we deliver in practice
Enterprise knowledge / RAG
Connect approved repositories, split content into searchable passages and preserve source metadata. Retrieval-augmented generation supplies relevant context at query time; it does not require training a model on every document.
Indexed knowledge set and refresh/removal workflow.
Permission-aware answers
Carry source permissions into retrieval and test user/group boundaries. Provide citations and a clear no-answer path when evidence is missing, conflicting or outside the visitor role.
Access-filtered retrieval and citation checks.
Tool-using agents
Expose only approved tool functions with validated arguments and restricted credentials. Separate read-only lookups from write actions; require confirmation for agreed higher-impact changes and log the outcome.
Tool catalogue, action policies and audit trail.
Human handover & evaluation
Define takeover triggers, transcript visibility and who resumes automation. Test unsupported questions, prompt injection attempts, stale knowledge and tool failures against an agreed evaluation set.
Agent evaluation report and handover runbook.
The technology, explained
RAG
Retrieval supplies relevant approved passages to a model at answer time. Grounding improves traceability but still requires permissions, citation checks and evaluation.
Embeddings & hybrid search
Embeddings represent semantic similarity. Hybrid search combines meaning-based retrieval with keyword matching to improve coverage of names, codes and natural-language questions.
AI agent vs chatbot
A chatbot primarily responds; an agent may select tools and perform bounded actions. Tool permissions and action approval matter as much as response quality.
Reference architectures & design practices
Illustrative patterns for project discussions. Final topology, sizing, products and acceptance criteria are developed against your requirements. These are BrightGate concept diagrams, not vendor-certified designs.
Permission-aware knowledge assistant+
On smaller screens, scroll the diagram sideways to see all components.
The query retrieves only content the user can access, and the model builds an answer from those passages. Sources remain visible. Separate ingestion updates the index when documents change or are removed; tool actions use an additional approval boundary.
Design practices to apply
- Propagate permissions before retrieval, not after generation.
- Treat retrieved documents as data, not instructions.
- Evaluate answer support and escalation quality.
Systems, products & official references
Examples to evaluate during scoping. Selection, supported features, hardware and licences depend on the agreed solution; a product mention does not imply a certification or an included purchase.
Hybrid / vector search
Search components for relevant passages, with document-level access filtering.
Official reference Microsoft retrieval-augmented generationAgent orchestration
Coordinate retrieval and approved tools; use an agent only when task complexity justifies it.
Official reference Microsoft agentic RAG architectureEnterprise repositories & SSO
Supported connectors and identity integration preserve customer ownership and access rules.
A result your team can accept
Acceptance evidence
Grounding
Review answer correctness and source support on representative questions.
Permissions
Attempt cross-user retrieval and verify restricted content stays inaccessible.
Actions
Test approval, duplicate submissions, failed tools and human takeover.
View the complete capability checklist
- Knowledge assistants with retrieval-augmented generation (RAG)
- Customer-service chat and agent handover
- IT service-desk and HR policy assistants
- Sales, proposal and research copilots
- Workflow agents for approved multi-step tasks
- CRM, ERP and ticketing tool integrations
- Permission-aware search and cited answers
- Agent evaluation, audit trails and human approvals