Building AI-powered features requires a growing vocabulary. These exercises cover the collocations ML engineers and AI product teams use when integrating LLMs, prompt engineering, and managing model outputs.
0 / 5 completed
1 / 5
The product team decided to ___ to power the new search feature rather than build a custom model.
Integrate an LLM is the standard software engineering collocation for embedding a large language model into a product or system. 'Integrate' implies connecting it with the existing architecture. 'Add' and 'use' are too vague; 'plug in' is informal and oversimplifies the effort.
2 / 5
The ML engineer spent a week iterating on the system prompt to ___ for the summarisation task.
Prompt engineer is now an established verb-phrase collocation in AI product development for the discipline of crafting effective prompts. Using it as a verb ('to prompt engineer') is standard in AI team vocabulary. 'Write' and 'create prompts' describe the activity but not the discipline.
3 / 5
Before launching the feature, the team needed to ___ across a diverse benchmark dataset.
Evaluate outputs is the standard AI/ML collocation for systematically assessing the quality and accuracy of model responses. 'Evaluate' implies structured, criteria-based assessment. 'Test' and 'review' are also used; 'check' is too informal for a formal evaluation pipeline.
4 / 5
The team added a fact-checking layer to ___ that could damage the product's credibility.
Handle hallucinations is the standard AI engineering collocation for designing systems that detect and mitigate model fabrications. 'Handle' implies building a systematic response to the problem. 'Prevent' implies elimination; 'avoid' is aspirational; 'fix' implies the model itself can be corrected.
5 / 5
The RAG architecture was chosen specifically because it allows the system to ___ efficiently during multi-turn conversations.
Manage context is the standard AI systems collocation for controlling what information is passed within the token window across conversation turns. 'Manage' implies active, intelligent orchestration of the context window. 'Handle' is close; 'store' implies persistence only; 'control' is too rigid.