Ontology-Driven Knowledge Engineering for Enterprise AI Office Agents

Knowledge Bottlenecks in Enterprise AI

Large enterprises possess massive, fragmented, and inconsistent knowledge assets—policies, contracts, technical documents, emails, and expert experience. Retrieval success rates are often below 40% and average search time exceeds 30 minutes, because knowledge is scattered across systems, terminology is inconsistent, and generated answers lack evidence, rules, and version control.

When large language models (LLMs) are plugged into office workflows, these problems are amplified: models can produce fluent text but cannot guarantee that the underlying knowledge is traceable, versioned, or compliant.

Three‑Layer Ontology‑Driven Architecture

The proposed architecture separates an enterprise office agent into:

Entrance layer : user‑facing interfaces (chat, search, workbench).

Middleware layer : LLM orchestration, RAG/GraphRAG pipelines, agent scheduling, permission control.

Knowledge‑engineering ontology layer : unified semantic model, rule constraints, evidence repository, version baseline, and expert feedback loop.

The real competitive moat resides in the ontology layer, which encodes business objects, processes, regulations, and evidence as machine‑executable, auditable structures.

Ontology Implementation Path

A six‑step modeling method converts business language into executable semantics:

Scenario definition – identify high‑value, high‑risk loops.

Term unification – resolve synonyms, aliases, and coding standards.

Entity modeling – define documents, clauses, devices, parameters, etc.

Relationship modeling – capture belongs‑to, applies‑to, cites, constrains, and evolves relations.

Rule modeling – express constraints with SHACL/SWRL (mandatory fields, value ranges, conflict detection).

Evaluation & iteration – measure recall, false‑positive rate, traceability, and update cycles.

The ontology consists of classes, attributes, individuals, and axioms that map directly to enterprise artifacts (e.g., a specific regulation clause becomes an individual with version, source, and applicability attributes).

Intelligent Office Applications

Three concrete use cases illustrate the approach:

Jing‑Review (璟审) : document compliance checking that returns a structured report with violated nodes, rule IDs, and remediation suggestions.

Jing‑Summarize (璟疏) : meeting‑minute generation enriched with speaker identity, role, agenda context, and traceable decisions.

Jing‑Sequence (璟序) : end‑to‑end document parsing, knowledge‑graph ingestion, governance, and assisted technical‑report generation.

All scenarios share a common pipeline: parse raw input, map to ontology entities, apply rule‑based validation, retrieve evidence via GraphRAG (vector + keyword + graph reasoning), and let the LLM generate output within the constrained evidence package.

Evaluation Metrics and Rollout Strategy

Six dimensions are used to assess the knowledge‑engineering effort:

Ontology quality – coverage, relationship completeness, synonym unification.

Retrieval quality – first‑hit success, evidence recall, path hit rate.

Review quality – miss rate, false‑positive rate, rule hit rate, human adoption.

Traceability – original‑text location rate, clause‑source completeness, version validity.

Iteration efficiency – rule‑update cycle, expert‑feedback loop, sample accumulation.

Implementation efficiency – delivery and deployment timeline.

A minimal viable ontology (MVO) can be built in roughly two weeks: days 1‑2 define the high‑value loop, days 3‑5 collect sample documents, days 6‑8 create entities and relationships, days 9‑10 author constraints, days 11‑12 run evaluation, and days 13‑14 integrate into the workbench.

Summary and Outlook

Ontology provides deterministic semantics, boundary enforcement, and evidence chains; LLMs supply natural‑language understanding and generation. Their collaboration, not substitution, is the key to production‑grade enterprise AI agents. Future work should continue to start from small, high‑frequency, measurable loops, gradually expanding the semantic graph, rules, and evaluation samples to form a sustainable AI knowledge base.