Ontology Semantics + Semantic Layers: Turning Enterprise Data into Reliable Business Answers

Why RAG Falls Short in High-Precision Scenarios

AI customer-service assistants are now standard on e-commerce platforms, yet they often stumble on questions that require combining multiple business rules. For example, when a user asks whether a specific item can be returned, the assistant may cite the platform’s generic “7-day no-reason return” policy, missing that the item was sold at a discount, is custom-made, or has special terms on its detail page. The root cause is not a lack of data but the loss of business relationships among data assets.

Four Core Problems with Naïve RAG

Chunk granularity: Return policies split across dozens of rules by channel, category, and promotion. Coarse chunks mix rules and drown signals; fine chunks break a single rule’s applicability context, leaving the model with half the information.

Fuzzy semantic matching only: RAG computes similarity over all words together. It cannot distinguish “brand-direct channel” from “third-party store” or “buy-3-get-1-free” from “discount-on-total,” so it retrieves return-related fragments but fails to apply the exact channel, category, promotion, and time constraints.

Broken reasoning chains: Determining return eligibility requires traversing order → product category → sales channel → promotion rule → applicable clause. A single vector lookup cannot follow these entity links; any missing link yields a related but inapplicable rule.

Low-frequency rules get buried: High-frequency rules like “7-day no-reason return” dominate similarity scores, while rare but correct rules (e.g., “brand-direct + restricted-category + mega-promo → only quality issues eligible for exchange”) rank too low to be retrieved.

Ontology Semantic Modeling: Restoring Business Logic

Volcano Engine’s Enterprise Knowledge Engine adopts a fundamentally different approach: instead of shredding documents, it reconstructs the structure and relationships among data assets so AI truly understands business logic. This is implemented in three steps:

Step 1: Identify Business Objects and Build a Relationship Network

The system extracts five core entities in the return scenario — Order, Product, Sales Channel, Promotion Batch, Return Rule — and links them via relational keys: which order belongs to which channel, contains which product category, participates in which promotion, and which combination of conditions triggers which return rule. This creates a relational map of the enterprise’s return knowledge.

Step 2: Multi-Condition Precise Matching Based on Business Semantics

When a user asks a natural-language question, the system first parses intent and identifies the four attributes of the current order (e.g., channel=brand-direct, category=beauty, promotion=buy-3-get-1-free, order-time=within promotion). It then automatically traverses the relationship network to lock onto the single rule that applies under that exact combination: “Promotional items only support exchange for quality issues after receipt.” No manual query orchestration is needed; the system decides and executes the traversal autonomously.

Step 3: End-to-End Traceability

Every answer cites the full provenance: rule text, rule ID, effective date, applicable channel and category scope, down to the original clause. Customer-service staff can verify each conclusion against its source.

Beyond E-Commerce: Aviation Maintenance Example

The same ontology approach works wherever answers require multi-hop business logic. In aviation maintenance, an engineer asking “How has the 1A engine’s collector-tank drain-pipe leak been handled historically?” triggers a chain: engine model → component → fault mode → repair procedure.

Verified Results

In high-precision scenarios, combining ontology semantics with RAG raised QA accuracy from 50–70% (RAG alone) to 100% , and cut problem-resolution time from 30+ minutes of manual lookup to minute-level response.

Semantic Layer: Giving AI a Deterministic Business Caliber

Return queries suffer from “retrieved but wrong.” A more pervasive enterprise pain is “vague questions, inconsistent answers” in daily business analysis. An analyst asking “How is our APP performing lately?” gets different answers each time: “lately” could be 7 or 30 days; “performance” could be WAU, retention, or DAU; “our APP” may include test-environment data. Humans fill these gaps with tribal knowledge; AI can only guess.

Five Composable Knowledge Types

The semantic layer systematizes the mapping from “human language” to “data” into five reusable categories (illustrated in the diagram):

Business terminology (synonyms, abbreviations)

Dimension enumerations (valid values for each dimension)

Atomic metrics (base measures with fixed formulas)

Derived metrics (compositions of atomic metrics)

Analysis frameworks (structured reasoning templates)

These five types converge into two complementary systems at consumption time:

Precise Data Retrieval — tells AI which table, which fields, and which caliber to use for calculation.

Business Understanding — tells AI which analytical angle to take and which framework to structure the conclusion.

Deterministic Four-Step Execution Path

With the semantic layer configured, the same “APP performance” question follows a fixed, repeatable path:

Intent Understanding: “Performance” maps to core operating metrics; “lately” defaults to “last 7 days.”

Caliber Locking: Semantic layer returns the product line’s core metrics: DAU + Day-1 Retention + Avg Session Duration.

Precise Data Fetching: Technical semantics locate the exact Hive tables and columns; deterministic SQL is executed.

Analysis Construction: Output is organized using the pre-defined analytical framework, delivering trend judgment and attribution suggestions.

Measured Improvement

In real enterprise analysis scenarios, after depositing business jargon, dimension enumerations, and atomic metrics into the semantic layer, QA accuracy reached 82.3% , significantly outperforming the no-semantic-layer baseline. Analysts no longer need repeated clarification; one question yields a caliber-correct result.

The essential value of the semantic layer is to make every query follow a deterministic path instead of relying on the model to guess.

Four Solution Combinations Covering the Full Enterprise Knowledge Spectrum

Beyond the two core capabilities, the Enterprise Knowledge Engine offers four deployable solutions that can be mixed per scenario (see diagram):

Ontology + RAG — covers precise multi-hop queries and general QA.

Semantic Layer + RAG — satisfies both analytical data retrieval and daily knowledge lookup.

Other combinations tailored to specific organizational needs.

Choosing the right combination for the scenario is key to unlocking the engine’s value.

Production Deployments Across Industries

Government services: QA system covering housing-fund, industrial-development, talent-attraction domains achieved a 93.75% issue-resolution rate.

Enterprise office collaboration: Stably serves 100,000+ documents and 200+ concurrent users; employees retrieve needed information via natural language.

Customer service: Assistant delivers second-level replies based on intelligent knowledge retrieval; manual ticket dispatch volume dropped sharply.

Only when data assets are not merely “stored” but can be accurately queried, correctly calculated, and fully traced do they truly work for the business.