AI glossary
$json
The json part of the item currently entering this node.
Acceptance Criteria
Specific, checkable conditions, written before running, that decide whether an output is accepted.
Accountability
Clarity about who answers for a decision and its consequences.
ACME
A protocol that automates getting and renewing TLS certificates by proving domain control.
Action Boundary
The line between what a system does itself and what it only proposes.
Adapter
Small code translating the program's common interface to one provider's API.
Agent Loop
The repeated cycle of choosing an action, running it, seeing the result and deciding the next step, until a stop condition is met.
Aggregate
Turning many items into one item holding their list.
AI Agent
A system in which the model decides its own steps and tool use.
AI Assistant
Direct human conversation with a model; the human decides the path each turn.
AI Disclosure
Telling the audience that content or a decision involved AI.
Alert
A message to an owner stating what failed, where, when and the next step, without sensitive data.
Alert Fatigue
When alerts are so many or so repetitive that people stop reading them and the important one gets lost.
Anonymization
Changing data by removal, substitution or generalization so a person is not identifiable.
API
An interface through which programs connect to a service; a successful response can still contain incorrect content.
API Key
A string issued by a service to identify requests as yours; valid until revoked.
Approval Gate
A point where a workflow pauses until a named human decides.
Approval Timeout
When the approver has not decided in time and the workflow follows a preset path.
Assistant Spec
A tool-independent document holding an assistant's instructions, knowledge list, I/O contract, tests and version log.
Attack Surface
All the ways someone can reach a system or make it act; every extra door is extra risk.
Audit Log
An immutable record of who did or approved what, and when.
Auto-fixing Output Parser
An output parser that asks the model to fix a failed parse, at the cost of an extra call.
Automated Sender
A message sent by a system (newsletter, auto-reply, bounce) that an email assistant must not answer.
Batch Processing
Sending non-urgent requests in batches, usually cheaper with slower results.
Batching
Processing items in small groups, usually pausing between groups to respect rate limits.
Bearer Token
A credential sent as Authorization: Bearer; whoever holds it is accepted.
Blast Radius
The maximum harm possible if a system is fooled or fails; depends on its access.
Bot Token
The key a platform issues for a bot; whoever has it can act as the bot.
Branch
A lightweight pointer to a commit that gives a separate line of change until it is merged.
Branching
Sending items down different paths based on conditions, with If or Switch.
Calibration
Agreement between expressed confidence and actual accuracy.
Channel-agnostic Design
Separating bot logic from platform details so switching channels is configuration only.
Chat Completions API
An endpoint that takes a list of messages and returns the model's reply with token usage.
Chat Memory
A session's message history, stored and re-sent to the model each turn.
Checklist
A short list of specific checks always done before delivery.
Claim
A statement that can be true or false; the smallest unit of verification.
Claimed Completion
When an agent says it did something but the actual tool result does not show it.
Classification
Choosing one label from a closed list for an input.
Commit
A snapshot of staged files with message, author and time that you can return to.
Community Node
A node written by someone other than n8n and installed on an instance; it runs with full access to the machine and workflow data.
Compose File
A YAML file defining several services, networks, volumes and their configuration for Docker Compose.
Configuration
Any value that differs between deployments (addresses, limits, keys); kept separate from code.
Constant-time Comparison
A comparison whose duration does not depend on where values differ, preventing timing attacks.
Constraint
A boundary the output must not cross, ideally stated with its reason.
Container
An isolated process run from an image; whatever is written to its writable layer is lost when it is removed.
Context
Everything sent with a request: instructions, conversation history, documents; the model's working memory.
Context Rot
Degradation of accuracy and recall as the number of tokens in context grows.
Context Selection
Deciding which information goes into the prompt for a task, using four filters: necessary, authoritative, fresh, permitted.
Context window
The maximum tokens a model considers in one request, including its output; overflow is not always reported.
Conversation History
Earlier messages resent to the model on every turn.
Credential
Where n8n stores keys or logins encrypted; nodes only reference it.
Data Mapping
Transforming one node's data into the shape the next node expects.
Data Minimization
Sending the least data needed, especially personal data, to do the task.
Data retention
How long and why a provider keeps inputs and outputs; not training is not the same as not retaining.
Deduplication
Dropping an input whose identifier was already processed, before any action repeats.
Delimiter
A marker such as <input>…</input> tags or ### that separates a prompt section, especially input data, from instructions.
Deterministic Step
A step that always gives the same output for the same input.
Direct Prompt Injection
A user's own message attempting to change the model's behavior.
DNS A Record
A DNS record pointing a name to a server's IPv4 address (AAAA for IPv6).
Docker Volume
Persistent storage outside a container that survives container removal.
Draft Guardrail
Code rules that check model output before it is shown or sent, e.g. rejecting discount or refund promises.
Duplicate Delivery
The same event arriving more than once; normal for many senders after errors or delays.
Duplicate Execution
Running the same work twice for one input so that an action happens twice.
Edge Case
An unusual input that raises the chance of failure: empty, very long, ambiguous or rule-breaking.
Edit Fields
An n8n node that creates, renames or removes fields.
Embedding model
Converts text to vectors to find semantically similar text; does not generate text.
Encryption Key
The key n8n uses to encrypt credentials in its database; backups of credentials are unreadable without it.
Endpoint
A specific API address that performs a specific operation.
Environment
A specific deployment of a program, such as dev, test or prod, with its own configuration values.
Environment Variable
A value set in the runtime environment outside the code and read by the program at run time.
Error Output
A separate node branch passing failed items with error details onward.
Error Workflow
A separate workflow starting with Error Trigger that runs when another workflow fails.
Evaluation set
Real samples of your task with scoring criteria, including edge cases, to compare options fairly.
Evaluator-Optimizer
A pattern where one model step generates and another evaluates against criteria and gives feedback.
Exact Match Grading
Automatic comparison of output to the correct answer; for closed answers.
Exception Path
The definition of cases outside the procedure and the role that handles them.
Execution
One run of a workflow, with each node's data and result recorded.
Execution path
How you run exercises: a legally available hosted service, a local model, or prepared sample runs.
Exponential Backoff
Increasing the wait exponentially after each failed attempt so the server can recover.
Expression
Short code inside {{ }} in a node field, evaluated per item at run time.
Extraction
Pulling specific fields out of free text.
- Structured Output: Tables, JSON and Schemas
Get output a program can read: tables and JSON, an output contract with JSON Schema, provider structured-output modes and three-level validation.
Fail-fast Validation
Checking all configuration at start-up and refusing to run when it is invalid, instead of failing later.
Fallback Branch
A path for items matching no condition so they are not silently lost.
Few-shot Prompting
Including a few input/output examples in the prompt so the model picks up the pattern.
Field Path
The address of a nested value, such as data[0].items[0].title.
Fine-tuning and RLHF
Steps turning a pretrained model into an instruction-following assistant; they do not remove errors and can encourage agreement.
Firewall
Rules deciding which incoming connections reach which ports; Docker-published ports may bypass it.
Flattening
Turning nested data into table rows with a clear rule for each column's level.
General level (B1–B5)
A description of general AI capability across six competencies, assigned from assessment evidence, not lesson count.
Generation
Writing new text such as a reply draft.
GPU and VRAM
The graphics card and its memory; a local model that fits in VRAM runs fast, otherwise it falls back to CPU and RAM and slows down.
Grounding
Basing an answer on specific text in context, usually with quotes.
Hallucination
A plausible but false statement produced by a model.
Header Auth
A generic credential adding a named header and value to each request.
Healthcheck
A periodic command or request showing a service is actually ready, not just running.
Heartbeat Monitoring
A scheduled job records 'alive' after each successful run, and an outside watcher alerts when that record gets old.
History Window
The number of most recent turns sent to the model.
HMAC Signature
A code computed from the payload with a shared secret so the receiver can verify origin and integrity.
Hosted model
A model run on the provider's servers, subject to its terms and data policy.
HTTP Header
A name-value pair carrying message metadata, such as content type or credentials.
HTTP Method
The verb stating the purpose of a request, such as GET to read and POST to create.
HTTP Request
The message a client sends: method, URL, headers and sometimes a body.
HTTP Response
The message a server returns: status code, headers and usually a body.
Human in the Loop
A design where a named human, with time, criteria and authority to reject, approves output before it takes effect.
Human Queue
Where uncertain or out-of-scope items go, with a clear reason, for a person to handle.
Human Review
Provider staff reading some submitted data, e.g. for safety or product improvement.
Idempotency Key
A unique ID per logical operation sent with a side-effecting request so the server can detect retries.
Idempotent
A property where repeating an identical request has the same server-side effect as doing it once.
Ignore Rules
Patterns in .gitignore that keep secrets, local data and build output out of Git's suggestions.
Image
An immutable bundle of the files, libraries and settings a service needs; changes mean building a new image.
Import cURL
Building HTTP Request node settings from a curl command; overwrites existing settings.
Indirect Prompt Injection
Instructions hidden in external content (web, email, files, knowledge-base documents) that influence the model when it reads them.
Inference
Using a trained model to generate output; the model does not learn from your conversation.
Input and output tokens
Tokens sent to and generated by the model; usually priced separately, output costing more.
Input Data
The material the model works on; it is data, not instructions.
Instruction
The part of a prompt that states what to do and under which rules.
Instruction Hierarchy
The rule for which layer's instruction wins in a conflict; a trained preference, not a security boundary.
Intermediate Output
The output of a chain step, checked before the next step.
Item
The unit of data passed between n8n nodes; each node takes and returns a list of items.
Item Linking
n8n's record of which earlier item each item was produced from.
Jailbreak
An injection aimed at making the model ignore its safety rules.
Jitter
Adding randomness to retry delays so clients do not retry in lockstep.
JSON
A standard text format for data exchange with objects, arrays, strings, numbers, booleans and null.
JSON Array
An ordered list of values in brackets, indexed from zero.
JSON Mode
A setting that guarantees valid JSON only, not schema conformance.
- Structured Output: Tables, JSON and Schemas
Get output a program can read: tables and JSON, an output contract with JSON Schema, provider structured-output modes and three-level validation.
JSON Object
A set of name-value pairs in braces; names should be unique.
JSON Schema
A standard vocabulary for describing the allowed shape of JSON data: field types, required fields and allowed values.
- Structured Output: Tables, JSON and Schemas
Get output a program can read: tables and JSON, an output contract with JSON Schema, provider structured-output modes and three-level validation.
Key Rotation
Regularly replacing a key with a new one and revoking the old.
Knowledge cutoff
The date up to which a model's knowledge is most reliable, usually earlier than the training data end.
Lab notebook
A table recording each experiment so it can be repeated and compared.
Latency
Time from sending a request to receiving the response.
Lateral Reading
Checking a claim or source by consulting other sources rather than trusting the text's appearance.
Learning Path
An ordered set of courses toward a larger goal; everyone starts with the foundations path.
Least Privilege
Granting only the access needed for the task, nothing more.
License (model)
Terms for using an open model; read the license text, not only its name.
Local Inference
Running a model on your own computer or server, without sending text to an external provider.
Local model
A model running on your own hardware; text is not sent to an external server, but quality is not on par with the best hosted models.
Long Polling
A request the server holds open until new data arrives or a timeout.
Lost in the middle
Accuracy drops when key information sits in the middle of a long context.
Mail Loop
Two automated systems endlessly replying to each other.
Manual Trigger
A trigger fired from the editor's run button; for building and testing.
Max output
The maximum response length, separate from the context window.
Merge
Combining two item streams into one, appended or matched on a field.
Message Roles
The label of each message in a conversation: system or developer, user and assistant.
Metrics Endpoint
An address that exposes a system's counters and gauges in a standard text format for monitoring tools.
Model Grader
An LLM that scores qualitative properties with a rubric, preferably a different model from the generator.
Model Tiering
Small models for simple steps, large only for hard ones, each validated.
Monitoring
Regular review of run outcomes by a named owner, with alerts for unusual patterns.
MTok (million tokens)
Common pricing unit; cost = tokens × price ÷ 1,000,000.
n8n
A visual node-based workflow automation tool available as a cloud service or self-hosted.
Nested Data
An object or array inside another object or array.
Net Time Saving
Time actually freed after subtracting review, correction, setup and maintenance time.
Next-token prediction
The core loop of a language model: score continuations, pick a token, append, repeat.
Node
One step of an n8n workflow, such as a trigger, HTTP request, condition or data change.
Non-retryable Error
An error that retrying will not fix, such as an invalid key or incomplete input.
OAuth
A way for a user to grant an app limited access on the service's own page without sharing a password.
Offset
An ID telling the platform not to send earlier updates again.
Open-weight model
A model with published weights that can be self-run, still subject to its license.
OpenAI-compatible Endpoint
Another provider's endpoint that mimics the OpenAI API shape, usually with differences.
Output Contract
An explicit agreement between the prompt and the consuming program about fields, types and meaning.
- Structured Output: Tables, JSON and Schemas
Get output a program can read: tables and JSON, an output contract with JSON Schema, provider structured-output modes and three-level validation.
Output Format
The expected shape, length and structure of the response, ideally checkable.
- Structured Output: Tables, JSON and Schemas
Get output a program can read: tables and JSON, an output contract with JSON Schema, provider structured-output modes and three-level validation.
Output Parser
An n8n sub-node that requests a format from the model and converts the reply into structure.
p95 Latency
The latency 95% of requests are faster than.
Pagination
Reading large data in successive requests, page by page, until a stop condition.
Parse Failure
When a model reply does not match the requested format and cannot be parsed.
Payload
The body of a request, especially a webhook, carrying event data.
Payload Validation
Checking the shape and values of incoming data before any action and rejecting invalid input explicitly.
Personal Data
Information relating to an identified or identifiable person.
Pinned Data
A fixed output kept on a node so later nodes can be tested without re-running it.
Polling
Regularly asking a service whether anything new has happened.
Predefined Credential Type
A credential type n8n provides for a specific service.
Pretraining
Initial training on a large text corpus to predict the next word.
Primary Source
Where a fact originates, such as an official document or original report, not a retelling.
Procedure Owner
The role that maintains a procedure, approves changes and answers questions.
Processing Ledger
A table recording what happened to each input id so nothing is processed twice.
Prompt
The text given to a model to perform a task, including instructions, context and input data.
Prompt Caching
Storing a fixed prompt prefix at the provider so re-reading it costs less.
Prompt Chaining
Running prompts in sequence where each output feeds the next.
Prompt Injection
Input that alters a model's behavior or output in unintended ways; direct or indirect.
Prompt Template
A fixed, tested prompt with slots for input, run by a human.
Provider abstraction
A layer between your work and the model that makes switching providers a configuration change.
Pseudonymization
Replacing identifiers with labels that can be reversed using a separate table.
Publish
Making a saved workflow version live for production runs in n8n 2.x.
Pull Request
A proposal to merge a branch that is reviewed and discussed before merging.
Quality Criteria
Measurable properties of acceptable output, each with a threshold and a way to measure it.
Rate Limit
A service's cap on requests or tokens per time window; exceeding it usually returns 429.
Readiness Check
A check that answers 'ready to take work', not just 'alive'; in n8n, the database is connected and migrated.
Reasoning Model
A model that generates reasoning tokens before its final answer; better at multistep work, slower and usually costlier.
Reasoning tokens
Tokens a reasoning model produces before answering; usually hidden but billed as output.
Reconciliation
Checking the real state of an external system before repeating an action that may or may not have happened.
Regression
A case that passed in the previous version and fails in the new one.
Repository
A project folder together with its full Git history.
Response Format
A parameter requesting JSON or schema-conforming output, if the provider enforces it.
- Structured Output: Tables, JSON and Schemas
Get output a program can read: tables and JSON, an output contract with JSON Schema, provider structured-output modes and three-level validation.
Restore Test
Restoring a backup on a separate instance and checking that data and secrets are really usable.
Retention Period
How long data is kept before deletion.
Retry On Fail
An n8n node setting that re-runs the node after a failure.
Retry-After
A response header saying how long to wait before retrying.
Retryable Error
An error likely to resolve on retry after a short wait, such as a temporary outage.
Reverse Proxy
A service that receives outside requests, usually terminates TLS and forwards them to internal services.
Reversibility
How easily and cheaply a wrong outcome can be undone.
Risk-proportional verification
The harder an error is to reverse, the stricter the check before acting.
Role Prompting
Assigning a role for the model to answer in; it changes tone and focus, not knowledge or authority.
Rollback Plan
Pre-written steps to return to the previous healthy state after a failed change or upgrade.
Rule-based Task
Work done with clear, repeatable rules rather than fresh judgment each time.
Run Evidence
Outputs, event logs and counts showing a system and its failure modes were actually tested.
Safe Stop
When a requirement fails, the assistant withholds its main output and states exactly what is missing instead of guessing.
Sampling and temperature
Controls sampling randomness; even at zero, identical outputs are not guaranteed.
Schedule Trigger
A trigger that runs a workflow at set times.
Schema Validation
Automatic check that data conforms to a schema.
Schema-in-Prompt
Putting the output format in the prompt instead of provider enforcement; no guarantee.
- Structured Output: Tables, JSON and Schemas
Get output a program can read: tables and JSON, an output contract with JSON Schema, provider structured-output modes and three-level validation.
Secret
Any value that lets someone act as you: a key, token, password or signing key.
Secret Manager
A service that stores secrets encrypted, restricts and logs access and simplifies rotation.
Security Audit
A periodic report of unused credentials, risky nodes, unprotected entry points and settings; n8n audit in n8n.
Self-check
A step asking the model to verify its answer against criteria before finishing; not a substitute for external checks.
Self-hosting
Running software on your own machine or server, owning install, security, backups and updates.
Semantic Validation
Checking that values are correct and supported, beyond shape; e.g. that an order ID appears in the input.
Sensitive Data
Data whose disclosure causes serious harm: health, financial and card data, children's data, confidential documents.
Session ID
An identifier separating one conversation's messages from others; comes from the channel, not user text.
Session Isolation
Ensuring no session sees another session's messages.
Simplest Sufficient Solution
The least complexity that meets the success criteria.
Source Metadata
Labels such as title, date and owner attached to each document so its origin is clear.
Specialist badge
Three independent badges for building capability, separate from the general level, requiring execution evidence and review.
SSH Key Pair
A private key on your computer and a public key on the server: login without sending a password.
Stable Instructions
Standing text present in every conversation of an assistant, stating mission, scope, style, red lines, examples and fallback behavior.
Staging Area
Where you choose exactly what goes into the next commit; git add stores the file as it is at that moment.
Standard Procedure
A written description of a recurring task: inputs, steps, output, checks, exceptions and owner.
State Summary
A short note of facts, decisions and open items carried forward instead of the full history.
Status Code
A three-digit result code; the first digit is its class: 2xx success, 4xx client error, 5xx server error.
Step Budget
The maximum number of steps, tool calls or cost an agent run may use.
Step Ledger
A table recording, per event, which external steps are done, in progress or pending, so a rerun resumes correctly.
Step Zero
Checking input before any work: required columns, units, summary and incomplete rows; stop and ask if a requirement fails.
Stop And Error
A node that deliberately fails an execution to trigger the error workflow.
Stop Condition
A rule written in advance that says when and in what state an agent run ends: success, partial or immediate stop.
Structured Output
Machine-readable output such as JSON or a table that can be validated against a schema.
Structured Output Parser
An output parser converting output to a JSON object per a schema.
Sub-workflow
A workflow called by another workflow, with defined input and output.
Summarization
Shortening text while keeping key points.
Sustainable Use License
n8n's main license allowing internal, personal and consulting use but not offering n8n to others to build workflows.
Sycophancy
A tendency to agree with the user even when wrong; tested with neutral questions and pushback.
System Prompt
The builder's fixed message for a conversation or system, holding rules, role and examples.
Task Decomposition
Breaking complex work into steps with defined input, output and check.
Task Inventory
A table of a role's repetitive tasks with frequency, time and risk attributes.
Task Runner
A separate process that runs Code node code so user code does not run inside n8n itself; external mode in production.
Terms of service and supported regions
Rules defining who may use a service, where and how; a hard filter before comparing quality.
Test Set
Fixed inputs with expected answers or criteria used to measure each version consistently.
Timeout
The maximum time to wait for a response; after it the outcome is unknown.
TLS Certificate
A document proving the server controls the domain, used to encrypt traffic.
Token
The smallest unit a model processes and bills; counts depend on language and tokenizer.
Token Usage
Input and output token counts reported in a response; the basis for cost.
Tokenizer
The rule that splits text into tokens; changing it can change the cost of the same text substantially.
Total Cost of a Step
Model cost plus retries, human review and fixed infrastructure cost for one workflow step.
Trigger
The event that starts a workflow run, such as an incoming email, a schedule or a new row.
Untrusted Content
Any text you did not write that enters the context; it may contain instruction-like sentences.
Update
An event a messaging platform sends to a bot, with a numeric ID.
Upsert
A write that updates the record with the same key or creates it; repeating it never creates a second record.
Usage
Response fields reporting actual input and output tokens; the right basis for cost estimates.
Value-Risk Matrix
Placing opportunities on value and risk axes to set the order of work.
Verification Log
A table of claims with checked source, status and action.
Verification Status
The outcome label of a claim check: verified, false, unverifiable or needs internal confirmation.
Version Log
A record of each prompt or model change, its reason and its result on the test set.
Version Pinning
Running an exact version instead of a moving tag like latest, so upgrades happen only by decision.
Version-bound Approval
Approval valid only for the exact version shown to the approver.
Webhook
An HTTP request a service sends to your URL when an event happens.
Webhook Authentication
A condition a webhook request must meet (secret header, JWT or HMAC signature) before the workflow runs.
Workflow
A predefined path of steps that runs automatically; the builder decides the path.
Workflow Versioning
Keeping every version of a workflow elsewhere (for example Git) so changes can be seen, reviewed and undone.
Zero-shot Prompting
A prompt with instructions only, no examples.