Writing ·
Injecting company jargon for generative AI
LLMs don't know a company's internal vocabulary, so they guess. A simple way to teach it to them from a YAML file, tested across model sizes.
Every company has its own vocabulary: team names, project phases, acronyms, internal product names. Large language models have never seen it, so when a request uses that jargon they fill the gap with a plausible guess, and that guess is often a hallucination.
In this article for SDG Group’s Medium publication I show how to inject that knowledge from a YAML glossary straight into the prompt. The worked example is an app that turns natural-language requests into valid arXiv API queries, and I compare how well the approach works across model sizes.
The code is on GitHub: ordovas/context-inject-arxiv.