Acceldata
AIO

Last updated: Oct 07, 2026 16:33 UTC

This page defines the terms used across the AIO documentation, with one definition for each term.

The terms are in alphabetical order. Where a term has its own page, the definition links to it.

Term Definition
Access key and secret key The credential pair your application uses to send telemetry through the gateway. The SDK sends them only when you set both. See Create an API key.
__REDACTED__ The value that replaces text matching a redaction pattern. It also replaces content that can't be parsed or couldn't be redacted. See Redact by pattern.
Check kind How a rule template decides. deterministic checks don't call a model. The other kinds are aggregate, judge, which uses a large language model (LLM) as a judge, and classifier, which calls a model that returns a class.
Collector endpoint The base URL of the OpenTelemetry Protocol (OTLP) collector that receives your spans, logs, and metrics over HTTP.
Column detail What the list of columns on a data reference means. listed means every column touched is listed. all means every column was read, as with SELECT *. partial means the statement named columns that couldn't be assigned to a table, so there may be more. empty means no column information.
Content capture Recording prompt, response, and message content on spans. It's on by default. To turn it off, set OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=false. When capture is off, the literal values in SQL statements on database spans are withheld too.
Contract revision The version of the telemetry contract that the SDK follows. The current revision is 1.0.0.
Conversation id The key that groups traces into a session. The SDK sets it as the span attribute gen_ai.conversation.id. It's empty when your code didn't set one. See Track users and sessions.
Cost The estimated cost of LLM calls in US dollars (USD). AIO adds it up only over LLM steps that have a price, and it has no value when none of them do. See Cost tracking.
Coverage The data quality state of an asset now, not at the time each trace read it. checked means a policy is enabled, scheduled, and ran in the last 24 hours. not_running means the asset has policies, but none of them is enabled, scheduled, and recent. not_checked means the asset is in the catalog but has no policies. not_in_catalog means AIO looked and found no asset. unknown means the match is pending or ambiguous, or the policies haven't been fetched yet. See Check quality coverage.
Data reference A pair of a namespace and a name that says which table, view, file, endpoint, or collection a span read or wrote, for example snowflake://acme-prod and SALES.PUBLIC.ORDERS. A data reference isn't an asset. AIO matches it to an asset in the Acceldata catalog, so a data source you register later still links to spans that are already stored. See Data quality overview.
Data source host map A map from the address your application connects to, such as localhost through a tunnel or a private-link alias, to the address the catalog has the data source registered under. It lets your application's queries still link to the asset. You set it with data_source_hosts or AIO_DATA_SOURCE_HOSTS. See Record the data your agent uses.
Digest A sha256:<hex> value that replaces a whole withheld value. A digest shows which value was in use, but not what it said. See Hide system prompts and tool definitions.
Direction Whether a span read or wrote the data in a data reference: read or write.
Endpoint The URL of the AIO server. The SDK uses it to find projects by name. Starting with SDK version 1.0.0.dev2, endpoint means the AIO server. Before that version, it meant the OTLP collector.
Environment The deployment environment your application runs in, set as deployment.environment. The default is default. A rule with no environment watches every environment.
Eval type One evaluation that AIO can run. Each eval type has an id, a label, a description, an enabled flag, a custom flag, and a threshold score.
Facet A dimension you can filter on in the filter rail of a list. See Explore traces.
Fail-open A guard that raises an error is treated as allow. This is the default for a guard pipeline.
Guard A check that runs inside your application on a prompt or a response. See Guard overview.
Guard action What a guard decides: allow, warn, redact, or deny, from least to most severe. When several guards run together in a pipeline, the most severe action wins.
Guard phase When a guard runs. preflight runs on the input before the model call. postflight runs on the output after the model call.
Input preview The user input for an operation, set as aio.input_preview. It's cut off at 2,000 characters.
Instrumentor A library integration that creates spans automatically. The SDK skips any integration whose library isn't installed.
Judge model A model that must be bound to the workspace before a rule from a template that needs a judge can run.
Match The result of matching a data reference to an asset in the catalog. exact is the only result that links to an asset. ambiguous is the second result. unresolved means the catalog has no such asset. pending means the catalog hasn't been asked yet, or couldn't be asked.
Mechanism The connector or SDK call that produced a data reference, for example sql_parse or declare_assets.
Model-input redaction scope A block of code, marked with redact_model_inputs, in which the input messages and system instructions of a model call are replaced by a digest. Output, and spans that aren't model calls, are left alone.
No value The facet label for rows where that dimension isn't set.
Node details event A gen_ai.agent.node.details log event that carries a LangGraph node's input and output. It's recorded only for nodes marked with @langgraph_node.
Offline evaluation A prompt and response pair that you send from the SDK to be scored against one or more eval types. See Run an eval.
Payload state How a step's content arrived: inline, truncated, redacted, or expired. It explains why a preview may be incomplete. Previews are taken after redaction and are a few KB in size.
Project The container your telemetry is sent to. See Create a project.
Project id A UUIDv7 (universally unique identifier, version 7) that identifies a project.
Project slug The part of a project's URL that names it. It's unique within a tenant and stays the same when you rename the project.
Project status The state of a project: Active, Archived, or Setup incomplete.
Quality coverage The project tab that shows which data assets your project's traces read or wrote, and whether anything checks them. It shows policy counts, state, and checked columns, but never policy names. See Check quality coverage.
Redaction policy The set of rules that withhold content before it leaves your application. Every switch is off by default. See Redaction overview.
Reference origin How AIO learned about a data reference. auto means a connector derived it or it was taken from the SQL. declared means your code stated it with declare_assets.
Release The version of your application, taken from service.version. When that isn't set, AIO uses the agent version hash.
Root span The span that root_span() opens for one unit of work, also called an operation. Its identity values are added to every span and log record created inside it.
Rule A saved check that belongs to one project. A rule has a template, name, description, enabled flag, environment, unit, severity, score type, sample rate, and parameters. Editing a rule creates a new rule version. See Rules overview.
Rule template What a rule is created from, for example banned_term or empty_output. Each template belongs to a category, such as Output integrity. See Create a rule.
Run id The id that groups the evaluations of one batch. The default is batch_ followed by 12 hexadecimal characters.
Sample rate The fraction of eligible traffic that a rule checks, from 0 to 1.
Score An evaluation result attached to a span as a gen_ai.evaluation.result event.
Score type The kind of score a rule produces: bool or numeric.
SDK name The instrumentation that produced a span.
Service The name of your instrumented application, set as service.name. The default is unknown_service.
Session The traces that share one conversation id. A session is also called a conversation. AIO assumes one user per session. See Track users and sessions.
Session status Whether a session had errors, shown as Has errors. A session is error when any span in any of its turns failed, and ok otherwise. The status sticks, so it isn't the outcome of the last message.
Severity How serious a finding is: LOW, MEDIUM, HIGH, or CRITICAL. When no severity is set, the UI shows —.
Span One operation inside a trace. The UI shows spans on the Spans tab and calls each one a step in labels such as Step id, Parent step, and Steps. The kinds are llm, tool, agent, chain, http, mcp, and retriever. The status is ok, error, or unset.
SQL literal redaction Removing the literal values from SQL statements on database spans. String literals become '?' and numeric literals become ?. The statement itself stays, because data references come from the tables it names. See Record the data your agent uses.
Storage mode Where a project's data is kept: Acceldata hosted, Customer store, or Existing OTel. Customer store and Existing OTel require a dataplane. When no mode is set, the UI shows —.
Suppressed span A span that's never recorded because a redaction rule names it. Every span below it is dropped too. See Drop attributes and spans.
Tag A key:value string that labels spans, where both the key and the value must be nonblank. Tags come from three sources: set in code with the SDK, derived from a span attribute by an attribute tag rule in project settings, or added by hand afterwards. See Tag traces.
Tenant The organization that owns the telemetry. Every span carries it as aio.tenant_id, and telemetry doesn't start without it.
Threat detected event A gen_ai.agent.threat_detected event attached to a span.
Tokens Total tokens are input tokens plus output tokens, where input tokens include cached tokens. Billable tokens leave out cached tokens.
Trace One unit of work. Each trace is one row on the traces list. Its totals can still change while the trace is open. The status is ok or error. See Tracing overview.
Turn One trace within a session. Last turn shows the name of the root span of the most recent turn.
Unit What a rule evaluates: span, trace, session, or window.
User id The end user of your application, set as aio.user_id. See Track users and sessions.
Verdict The outcome of one rule on one entity it checks. VERDICT_PASS means the check passed. VERDICT_FAIL is a finding. VERDICT_ERROR means the rule couldn't decide. VERDICT_SKIPPED always comes with a reason for the skip.

Next steps