AI terms, in plain language.
Short definitions of the words this site uses. If you're new to the field, start here; if you're not, skip ahead.
- Agent
- A model that can plan steps and use tools — search, databases, APIs — to complete a task, not just answer once.
- BM25
- A classic keyword-search scoring method — good at exact terms, codes and names that meaning-based search can miss.
- Chunking
- Splitting documents into smaller pieces so they can be searched and fed to a model.
- Context window
- How much text a model can read at once. Anything beyond it has to be left out or summarised.
- De-identification
- Removing or replacing the details that identify a person, so the rest of the data can be used more safely.
- Deterministic
- Gives exactly the same output every time for the same input.
- Embedding
- A list of numbers that captures what a piece of text means, so similar meanings end up close together.
- Eval set
- A fixed set of questions with known good answers, used to check whether a change made the system better or worse.
- Fine-tuning
- Further training an existing model on your own examples so it behaves better for your task.
- Guardrails
- Checks placed around a model to stop unsafe input going in or unsafe output coming out.
- HIPAA Safe Harbor
- A US method for de-identifying health data by removing 18 specific kinds of identifier.
- Hybrid search
- Running keyword search and vector search together and merging the results.
- Inference
- Running a trained model to get an answer — as opposed to training it.
- LLM
- A large language model — software trained on huge amounts of text that can read and write language.
- LLM-as-judge
- Using a model to grade another model's answers against a rubric.
- LoRA / QLoRA
- Cheap fine-tuning methods that train a small add-on instead of the whole model.
- MCP
- Model Context Protocol — an open standard for connecting models to tools and data sources.
- OCR
- Optical character recognition — turning scanned pages and images into machine-readable text.
- PHI
- Protected health information: anything that can identify a patient alongside their health details.
- In the US, HIPAA governs how PHI is stored, shared and de-identified.
- PII
- Personally identifiable information — names, addresses, ID numbers and anything else that points to a person.
- Prompt injection
- Hiding instructions in text a model reads — a document, a web page — to make it do something it shouldn't.
- Quantization
- Storing a model's numbers with less precision so it needs less memory and runs faster, at a small cost in quality.
- RAG
- Retrieval-augmented generation: find the relevant passages first, then have the model answer using only those.
- It lets a model answer from your own documents instead of from memory, and makes it possible to cite where each answer came from.
- Reranker
- A second, more careful model that re-orders search results by how well they actually answer the question.
- Semantic cache
- Reusing a previous answer when a new question means the same thing, even if worded differently.
- Token
- The unit models read and bill by — roughly three-quarters of an English word.
- Tokenization (privacy)
- Swapping a sensitive value for a placeholder like [PATIENT_1], with the real value kept in a separate locked store.
- Vector search
- Finding text by meaning rather than exact words, by comparing embeddings.
Contact
Working on something
like this?
I'm open to AI engineering, architecture and training work. Tell me what you're building and what the constraints are — that's usually enough to start.
Prefer a short form? Send a project brief
- Taking on new projects
- Usually replies within a day