English for AI Engineers: Key Vocabulary

Artificial intelligence is one of the fastest-moving fields in technology — and its vocabulary is evolving just as quickly. For non-native English speakers working in AI and ML, knowing the right words is not just about understanding papers and documentation. It is about communicating credibly in design meetings, code reviews, and product discussions.

This guide covers the core vocabulary you need, with definitions, example sentences, and tips on how each term is used in context.

Foundations: How Models Work

Embedding

An embedding is a numerical representation of data (text, images, audio) in a high-dimensional vector space. Items with similar meaning are positioned close together in that space.

“We store document embeddings in a vector database so we can retrieve semantically similar chunks at query time.”

“The model maps each token to a 1,536-dimensional embedding before processing.”

Inference

Inference is the process of using a trained model to make predictions on new data. It is the opposite of training.

“Inference latency is a bottleneck — we’re seeing P99 response times above two seconds.”

“We run inference on GPU instances to meet our throughput requirements.”

Token

A token is the basic unit of text that a language model processes — roughly a word or word fragment.

“This prompt uses around 800 tokens, which keeps us well within the model’s context window.”

Training and Adaptation

Fine-tuning

Fine-tuning is the process of taking a pre-trained model and continuing to train it on a smaller, domain-specific dataset to adapt its behaviour.

“We fine-tuned the base model on our support ticket data to improve classification accuracy.”

“Fine-tuning requires significantly less compute than training from scratch.”

Pre-training

Pre-training is the initial phase where a model learns from a large general dataset before any task-specific adaptation.

“The model was pre-trained on a trillion tokens of internet text.”

RLHF (Reinforcement Learning from Human Feedback)

RLHF is a training technique where human raters score model outputs, and those scores guide further training to align the model with human preferences.

“RLHF is how most commercial chat models are aligned to follow instructions helpfully.”

Retrieval and Context

RAG (Retrieval-Augmented Generation)

RAG is an architecture where a language model’s responses are augmented by retrieving relevant documents from an external knowledge base at query time.

“We use a RAG pipeline so the model always answers based on up-to-date company documentation, not just its training data.”

“The retrieval step returns the top five document chunks, which we inject into the prompt context.”

Context Window

The context window is the maximum amount of text (measured in tokens) a model can process in a single call.

“The 128k context window lets us pass the entire codebase into the prompt for refactoring tasks.”

Chunking

Chunking is the process of splitting large documents into smaller pieces before embedding them for retrieval.

“We chunk documents at 512 tokens with a 50-token overlap to preserve context across boundaries.”

Agents and Orchestration

Agent

In AI, an agent is a system that uses a language model to reason, plan, and take actions (such as calling tools or APIs) to accomplish a goal over multiple steps.

“The agent autonomously calls the search API, reads the results, and synthesises an answer.”

“We built a coding agent that can read files, write patches, and run tests in a sandboxed environment.”

Tool Use / Function Calling

Tool use (also called function calling) is a capability that allows a language model to invoke external functions or APIs as part of generating a response.

“We expose a weather API as a tool, and the model decides when to call it based on the user’s question.”

Orchestration

Orchestration refers to coordinating multiple AI components, agents, or model calls in a defined workflow.

“We use LangGraph for orchestration — it manages the state machine that routes between the planner and executor agents.”

Evaluation and Quality

Hallucination

Hallucination is when a model generates plausible-sounding but factually incorrect information.

“Our evaluation pipeline flags hallucinations by cross-referencing model outputs against the source documents.”

Benchmark

A benchmark is a standardised test used to measure model performance on a specific task or capability.

“The model scores 87% on the MMLU benchmark, which covers multi-domain knowledge.”

Latency vs. Throughput

Latency is the time to complete a single request. Throughput is how many requests the system can handle per unit of time.

“We optimised for latency in the customer-facing product, but we batch background jobs to maximise throughput.”

Putting It Together: Useful Phrases

• “The inference pipeline processes requests with a median latency of 340ms.”
• “We’re evaluating whether fine-tuning or RAG gives better accuracy for this use case.”
• “The agent uses tool calls to look up real-time data before generating the response.”
• “Chunking strategy has a significant effect on retrieval quality — we tested three approaches.”

Fluency with this vocabulary will help you participate confidently in AI engineering discussions, write clear technical specifications, and communicate progress to stakeholders. The field evolves rapidly, but these core terms form a stable foundation.

Navigating the Nuances of Feedback & Collaboration

For non-native speakers, understanding how to communicate technical ideas effectively is just as important as knowing the terminology itself. It’s not enough to simply translate; you need to grasp the cultural expectations around feedback, collaboration, and requesting assistance within a software development environment – particularly in fast-paced AI projects. A lot of the vocabulary we’ve covered – like “fine-tuning” or “inference” – can sound incredibly precise when spoken, but often requires context to be used appropriately.

One common scenario is receiving a code review comment. Let’s say you’ve submitted a pull request containing a new function designed to improve model accuracy using embeddings. A reviewer might leave a comment like: “This looks good, but could we explore adding some logging around the embedding retrieval process? It would really help with debugging if something goes wrong.” Simply translating “debugging” into your native language doesn’t fully convey the expectation here. The reviewer isn’t just pointing out a potential problem; they’re requesting information and potentially suggesting a proactive step to prevent future issues. Responding effectively requires more than a literal translation – it requires understanding that the comment is an invitation for clarification, not necessarily criticism. Framing your response as “Understood – I’ll add logging around the embedding retrieval. Could you provide some guidance on what level of detail would be most helpful?” demonstrates engagement and willingness to collaborate.

Similarly, Slack conversations can quickly become complex. Consider a situation where a team is discussing the performance of a Retrieval-Augmented Generation (RAG) pipeline. A developer might post: “The RAG latency is spiking during peak hours. We need to investigate potential bottlenecks in the vector database query.” This statement utilizes technical jargon, but it also subtly implies a responsibility – “we” need to investigate. It’s crucial to acknowledge this implied ownership and offer solutions, rather than simply stating the problem. A better response might be: “Okay, let’s start by profiling the database queries during peak hours. I can run some redis-cli commands to examine query times and identify any slow operations – something like redis-cli --stats could give us a good overview.”

Finally, when writing pull request descriptions, being specific about why you’re making changes is vital. Avoid vague statements like “Improved model performance.” Instead, articulate the impact: “Implemented fine-tuning on the sentiment analysis model using dataset X to improve accuracy by 2.5% as measured by F1-score.” This level of detail demonstrates a clear understanding of your work and its value.

redis-cli --stats | grep 'keyspace_hits'

This command, executed against a Redis instance (commonly used in RAG pipelines), allows you to monitor the number of successful key lookups – a critical metric for evaluating the efficiency of your vector database queries. Understanding and utilizing tools like this demonstrates technical proficiency beyond just knowing the vocabulary.