Glossary
Agent
A system that can read context, make decisions, and take actions to achieve a goal, often by calling tools and iterating on results without step-by-step human direction.
Read more →Anthropic
The company behind Claude. Claude offers three tiers of models (Haiku, Sonnet, and Opus), each suited to different kinds of work and with different costs per inference.
Chunking
The process of splitting a large document into smaller, overlapping pieces before embedding them. A 30-page baggage policy, for example, would be broken into paragraph-sized chunks so that a search can return just the relevant section rather than the entire document.
Context Window
The amount of text an LLM can consider in a single request, including your question, any retrieved documents, and the model's own reply. A larger context window means the model can work with more information at once, but costs more tokens.
Embedding
The process of converting text into a vector (a long list of numbers) that captures its meaning. Words and phrases with similar meanings end up with similar vectors, which is what allows a vector database to find related content even when the exact words differ.
Feedback Loop
A cycle where the output of a system is fed back as input, allowing it to self-correct. In agent design, this usually means the agent observes the result of an action and decides what to do next.
Gemini
Google's suite of LLM capabilities. It offers native integration with Google's search engine, Imagen for image generation, and is known for having an extremely long context window, meaning it can work with more of a conversation's history than some other LLMs.
Grounded
Constraining an LLM's answer to be based on specific retrieved sources or tool outputs (and ideally cite them), rather than relying on its general training knowledge.
Grounding
Constraining an LLM's answer to be based on specific retrieved sources or tool outputs (and ideally cite them), rather than relying on its general training knowledge.
Inference
A single request-and-response cycle with an LLM. When you ask a question and receive an answer, that is one inference. Providers typically charge per inference based on the number of tokens consumed.
LLM
Large Language Model. An AI model that interprets a request or question and generates an answer. LLMs are generative, meaning the answer is produced by the model's reasoning each time. ChatGPT, Claude, and Gemini are the most well-known examples.
MCP
Model Context Protocol. A standard for connecting AI models to external tools and data sources. Rather than each tool needing a custom integration, MCP provides a common interface so an AI model can query a loyalty database, check a booking system, or call an API through a single, consistent protocol.
Metadata
Data about data. For example, a chunk from the baggage policy might carry metadata like "Baggage Overview" and "Central Knowledge Base" to help with filtering and organisation during retrieval.
OpenAI
The company behind ChatGPT. It offers a wide range of capabilities including Codex (coding), Sora (video), and image generation. ChatGPT is available in multiple tiers, with "thinking" versions of its higher-tier models that consume more tokens but can produce more thorough answers.
Precision
Of the chunks returned by a query, how many are actually relevant? Did we avoid pulling in irrelevant results?
RAG
Retrieval-Augmented Generation. A pattern where the system retrieves relevant source passages and supplies them to the LLM so it can answer grounded in that text, often with citations. This means the model can reference up-to-date, domain-specific content (like policies and procedures) instead of relying on what it learned during training.
Read more →Recall
Of all the relevant chunks that exist, how many did we return? Did we miss anything?
Relational Database
A database where records (rows) are organised into tables, each with a consistent set of fields (columns). You can query across tables using SQL, e.g. "Find me all bookings for Passenger Sam in 2025." A strong fit for structured data.
Structured Data
Data that follows a consistent and predictable structure, like bookings, loyalty activity, and flight schedules. This data is naturally suited to databases and can be queried directly.
Token
The basic unit that an LLM reads and produces. A token is roughly a word or part of a word, for example "rebooking" might be split into "re" and "booking". LLMs have a cost per token and a maximum number of tokens they can handle per request.
Tool
An external capability that an agent can invoke, such as a database query, API call, or code execution environment. Tools let the agent interact with the world beyond its own reasoning.
Unstructured Data
Data that lives in documents or knowledge bases, like cancellation policies, operating procedures, or internal wikis. It can't easily be fitted into a traditional database due to its lack of a consistent structure.
Vector Database
A database that stores embedded chunks of text as vectors. Chunks with similar meaning end up close together, so when the LLM searches for "weather cancellations", chunks about rebooking, re-routing, and cancellation policy are all returned because they are semantically related.