04 / Enterprise knowledge and AI service
DEXIN AI
An assistant for enterprise knowledge and inquiries. This case focuses on permission-aware internal retrieval, cited answers, and human review.
GitHub · AI module documentation ↗The Challenge
Enterprise knowledge was dispersed across documents and tacit experience, making consistent, bounded answers difficult.
Design blueprint / KNOWLEDGE
The anatomy of an answer
Governed sources, permissions, and evidence shape the answer.
Skip exhibit, read decisions ↓
Concept model · not a simulationSources
Use governed sources with recorded versions and provenance.
ILLUSTRATIVE SCENARIOS
Three answer boundaries
Illustrative input: find a business process guide
Example: retrieve an accessible process document → return a summary with sources → invite review.
A citation is a starting point, not a verdict
These are proposed acceptance checks, not production evaluation results. Read-only access does not replace access control or human review.
Permission
Is restricted material excluded before context assembly, rather than merely hidden in the answer?
Support
Does the cited passage support the claim, with a compatible date and scope?
Gaps
Are missing or conflicting sources made explicit, with a next step for evidence or human review?
Read the full reasoning · Chinese →
Approach
- Organized sources by business topic
- Designed boundaries for retrieval, response, and human review
- Prioritized verifiable, high-frequency use cases
System Structure
- Permissions: the signed-in user's database access scope
- Retrieval: intent classification, keyword/database filtering, and deterministic ranking
- Answers: bounded context, source labels, and recent conversation
- Boundary: no creating, approving, paying, or editing business records
Design Decisions
Read-only first: permission and evidence precede the answer
Is a fluent answer enough to trigger a business action?
Separating inquiry from writes makes sourcing, permission, and human review inspectable. Restricted sources are not retrieved; missing evidence is stated explicitly. This exhibit explains the interaction without connecting company data or claiming verified operational reliability.
德小馨 AI · Public product scope ↗Outcomes
- Authenticated retrieval and answering endpoints implemented in the codebase
- Source-labelled answers and persisted recent conversations
- Business writes and autonomous execution remain outside the current scope
Discuss this project
Interested in this project’s methods, implementation, or tradeoffs? I welcome a conversation by email.