Case Study: Ontology Implementation Pitfalls in a Women’s Apparel Startup

The author discusses a conversation with a female developer who built an ontology for a women’s clothing private company, claiming that her analysis is accurate by using a static graph, MCP standard interfaces, and database queries to feed a large language model (LLM). The author points out that the core capability of an ontology should be complex inference, not just data extraction.

In the current setup, the static graph is used to locate nodes, which then trigger SQL queries to retrieve data. The retrieved data is handed to the LLM for analysis. This approach relies only on graph relationships, database queries, and LLM processing, without leveraging ontology reasoning to compute paths or values.

The intended ontology reasoning chain is described as entity → scenario → role → rule → action → entity. For example, the system should distinguish between a “first‑half‑year scenario” and a “second‑half‑year scenario” rather than letting the LLM infer time periods and query data on its own, which can lead to hallucinations.

Figures illustrate the workflow: Figure 1 shows the static graph used for data analysis; Figure 2 depicts a report generated after the LLM queries data via MCP; Figure 3 outlines the LLM’s call‑graph logic; Figure 4 defines the static node data; Figure 5 presents functions modeled as nodes; and Figure 6 visualizes the inference process where the LLM fetches two sales figures and assembles a comparison report. The author emphasizes that the LLM should only receive the final conclusion from the reasoning engine, not perform the intermediate calculations.

The author argues that the developer’s implementation resembles a React + graphRAG solution, lacking a solid, ontology‑driven inference chain. It is essentially a hard‑coded SQL analysis model, not a true ontology system that can autonomously filter and compute data based on scenarios.

In a proper ontology, the reasoning engine derives the necessary data by traversing multiple nodes, tables, rules, and hops, providing a fixed inference path. This prevents the LLM from generating hallucinations, as the reasoning engine handles the complex calculations while the LLM only interprets the final results.

Finally, the author thanks the netizen who provided the material and invites readers interested in ontology to join a discussion group.