GitHub Copilot に「オントロジー」を持たせたら目的によっては確かに減った。 Adding a domain ontology to GitHub Copilot reduced erroneous and off-target suggestions fo…

A developer experiment published on Qiita examined whether supplying a domain ontology to GitHub Copilot could reduce the frequency of incorrect or off-target code suggestions. The reported conclusion is deliberately qualified, but it is still noteworthy: for certain kinds of tasks, giving the assistant a structured description of domain concepts appears to have measurably lowered the rate of erroneous output. The result suggests that knowledge representation techniques, long associated with the semantic web and expert systems, may have renewed practical value in the era of AI coding assistants.

An ontology, in this setting, is a formal and structured description of the concepts within a particular domain together with the relationships among them. Instead of leaving the model to infer what terms such as "order," "customer," or "inventory item" mean purely from surrounding code and a natural-language prompt, an ontology states these entities explicitly, along with their attributes and how they relate to one another. The reasoning behind the experiment is that the large language models powering Copilot generate suggestions from statistical patterns learned during training, and that this pattern matching can drift when a project uses terminology in idiosyncratic, overloaded, or highly specialized ways. A clear definition of the domain vocabulary is intended to anchor the model's output to the intended meaning.

According to the write-up, the effect was not uniform across all situations. The improvement was most visible for specific, well-bounded purposes rather than as a blanket gain across every coding scenario, which is why the author frames the outcome cautiously. This pattern is consistent with how context injection generally behaves: extra structured information tends to help most when a task depends on domain semantics that the model would otherwise have to guess, and it offers less benefit for routine code where the model already performs well. The honesty of that framing is part of what makes the experiment worth reading, since it avoids presenting the technique as a universal fix.

Mechanically, the approach belongs to the broader family of prompt engineering and context-grounding strategies. Providing an ontology to Copilot is, in practice, a way of feeding curated knowledge into the model's context window so that suggestions are conditioned on it. This is conceptually adjacent to retrieval-augmented generation, where relevant documents or data are fetched and inserted into the prompt to improve factual grounding. It also connects to features that GitHub and other vendors have been building out, such as custom instructions, repository-level context, and Copilot's ability to reference workspace files. An ontology can be seen as a particularly disciplined version of that same idea: rather than ad hoc notes, it is a coherent model of the domain that the assistant can lean on.

The wider context helps explain why this is interesting. Accuracy and the tendency to produce plausible-but-wrong output remain central concerns for AI coding tools, and teams have explored several mitigations, including better prompting, fine-tuning, linting and test-based guardrails, and richer context retrieval. Reusing ontologies fits this landscape because the artifacts may already exist in some organizations, for example in data modeling, knowledge graphs, or schema documentation, meaning the marginal cost of trying the technique can be relatively low. It also revives interest in knowledge representation methods that predate the current generation of language models, hinting at a hybrid direction where formally structured knowledge complements statistical learning.

Several caveats apply. The findings come from an individual blog experiment rather than a controlled study, so the results should be read as suggestive rather than definitive, and they may not generalize to other languages, frameworks, or team workflows. Building and maintaining an ontology is itself an investment, and an outdated or inaccurate ontology could plausibly mislead the model as easily as a good one helps it. The size of the context window and how the ontology is formatted are also likely to influence results. Even so, the experiment offers a concrete, reproducible angle on a practical question: how to make AI coding assistants more reliable for domain-specific work. For developers wrestling with off-target suggestions in specialized codebases, structuring and supplying domain knowledge appears to be a low-risk technique worth evaluating on their own tasks.