GaplessGlossary
Definitions of the terms used across these docs, so every page uses each word the same way.
Each term has one meaning across all pages. If a page uses a word differently, that is a bug in the docs.
Core terms
| Term | Meaning |
|---|---|
| Assistant | The AI product described in these docs. Users chat with it, and it answers, summarizes, and takes actions |
| Tenant | A group whose data must be isolated from other groups, such as a customer company or a business unit |
| Principal | An identity that can be granted access: a user, a group, or a service account |
| ACL | Access control list. The list of principals allowed to read a document |
| Orchestrator | The service that runs the steps of a request: load history, retrieve, build the prompt, call the model, call tools |
| Model gateway | The internal service all model calls go through. Handles providers, retries, fallback, and cost tracking |
Models and prompts
| Term | Meaning |
|---|---|
| LLM | Large language model. A model that takes text in and produces text out |
| Token | The unit models read and write. Roughly 4 characters or three quarters of a word in English. Prices and limits are counted in tokens |
| Context window | The maximum number of tokens a model can read in one call, including the prompt and the answer |
| Prompt | Everything sent to the model in one call: instructions, retrieved content, history, and the question |
| System prompt | The instructions part of the prompt, written by the developers, not the user |
| Prompt template | A versioned prompt with placeholders that the orchestrator fills in for each request |
| Temperature | A setting that controls randomness. Low values give more consistent output |
| Time to first token (TTFT) | Delay between sending a request and receiving the first output token. The main speed metric users feel |
| Hallucination | A confident statement from the model that is not supported by its sources or by fact |
| Fine-tuning | Further training a model on your own examples to change its behavior on a task |
Retrieval
| Term | Meaning |
|---|---|
| RAG | Retrieval-augmented generation. Search for relevant content first, then give it to the model to answer from |
| Ingestion | The background process that turns source documents into searchable chunks |
| Connector | Code that pulls documents and their permissions from a source system, such as a wiki or a file drive |
| Chunk | A piece of a document, usually a few hundred tokens, that is indexed and retrieved as one unit |
| Embedding | A list of numbers that represents the meaning of a text. Similar meanings give nearby embeddings |
| Vector index | A data structure that finds the embeddings nearest to a query embedding |
| Keyword search | Search by matching words, usually ranked with BM25. Good for names, codes, and exact terms |
| Hybrid search | Combining keyword and vector search results |
| Re-ranking | A second, more accurate scoring of the top search results, usually with a cross-encoder model |
| Top-k | The number of results kept from a search step |
| Grounding | How well an answer is supported by the retrieved sources |
| Citation | A reference in the answer pointing to the source chunk that supports a statement |
Agents and tools
| Term | Meaning |
|---|---|
| Tool | A function the model can ask the system to run, such as search_tickets or create_ticket |
| Tool call | The model's structured request to run a tool, with arguments |
| Agent | A loop where the model decides which tools to call, sees the results, and continues until the task is done |
| Human in the loop | A step where a person must approve before the system continues, usually before a side effect |
| Side effect | Any action that changes something outside the assistant: writing, sending, deleting, paying |
Quality and safety
| Term | Meaning |
|---|---|
| Eval | A repeatable test that scores the system's outputs on a fixed dataset |
| Golden dataset | A curated set of inputs with known good answers or grading criteria, used for evals |
| LLM-as-judge | Using a model to grade another model's output against criteria |
| Regression | A change that makes a previously passing case fail |
| Guardrail | A check on input or output that blocks or modifies content that breaks a rule |
| Prompt injection | Text that tries to override the system's instructions, typed by a user or hidden in a document |
| Jailbreak | A prompt crafted to make the model ignore its safety rules |
| PII | Personally identifiable information, such as names, emails, phone numbers, and ID numbers |
| Red teaming | Deliberately attacking the system to find safety and security failures before users do |
Operations
| Term | Meaning |
|---|---|
| Trace | The record of one request across every service and step, with timings |
| Span | One step inside a trace, such as a retrieval call or a model call |
| SLO | Service level objective. A target such as "99.5% of messages get a first token within 3 seconds" |
| p50, p95, p99 | Percentiles. p95 latency is the value that 95% of requests are faster than |
| Canary | Releasing a change to a small share of traffic first, to catch problems before full rollout |
| Runbook | Step-by-step instructions for handling a known operational problem |