Filters

The Filter panel in the trace explorer narrows which traces are shown. It offers three modes: plain-language AI search, a Basic property/condition/value builder, and a Query expression for power users. This page lists the modes, every property you can filter on, the metrics you can filter and aggregate by, and a set of ready-to-paste queries.

Filter modes

Mode	Use it for
AI search	Describe what you want in plain English (e.g. “errors on gpt-4o today”) and the filter is built for you
Basic	Pick a property, a condition, and a value. Add several; they apply together (AND)
Query	Write a filter expression directly, for conditions the Basic builder can’t express

Filterable properties

Each property maps to a span attribute key. Use the property name in the Basic builder, or the attribute key in a Query expression.

Property	Attribute key	Example value	Operators
Trace ID	trace.id	7f3c1a9b…	is, is not
Trace Name	trace.name	support_agent.run	is, is not, contains
Span Name	span.name	tool.check_order_status	is, is not, contains
Status	status	OK, ERROR	is, is not
Model	llm.model_name	gpt-4o	is, is not, contains
Node Type	node.type	llm, chain, tool	is, is not
Span Kind	fi.span.kind	LLM, RETRIEVER, TOOL	is, is not
User ID	user.id	user_8821	is, is not, contains
Provider	llm.provider	openai, anthropic	is, is not
Service / Trace Name	service.name	checkout-service	is, is not, contains
Latency	latency	5000 (ms, numeric)	greater than, less than
Eval score	eval.score / scores.<name>	0.5 (numeric)	greater than, less than
Tag	tag.tags	needs-review (list of strings)	contains

In addition, annotation values attached to a span are filterable, using is / is not on an annotation value.

Ready-to-use filters

Paste any of these into the Query tab. Each line finds a common class of trace.

status = ERROR AND llm.model_name = gpt-4o    # errors on a specific model
latency > 5000                                # slow spans, over 5 seconds
user.id = user_8821                           # every trace for one end user
eval.score < 0.5                              # low-scoring responses
fi.span.kind = RETRIEVER                      # retriever spans only
tag.tags contains needs-review               # traces carrying a tag

Basic operators

Operator	Applies to
is / is not	Exact match (status, model, provider, enums)
contains	Substring match (names, inputs, user ID)
greater than / less than	Numeric values (latency, tokens, eval score)

Metrics

Alongside the properties above, these are the metrics Observe computes from your spans: the values you sort the trace table by and aggregate on a dashboard widget. Nothing is precomputed, each is derived from the spans that match your filters and time window.

Metric	Unit	What it measures
Span count	count	Number of spans matching the filters
Error count	count	Number of spans or traces with ERROR status
Span response time	ms	Latency of a span
LLM response time	ms	Latency of LLM spans specifically
Token usage	tokens	Tokens consumed (prompt + completion)
Cost	USD	Computed cost of model calls
Eval pass-rate	%	Share of evaluated spans that passed their eval

• Latency is per span, not per trace. Span response time measures one operation; a whole request’s wall-clock time is the root span’s duration, so filter to root spans to compare requests.
• Cost and token usage only populate on LLM spans. A tool or retrieval span adds to span count but contributes zero tokens and zero cost.

Aggregations and granularity

Aggregate a metric with Sum, Average, Median, Count, Distinct count, or Min / Max, over a time bucket of minute, hour, day, week, or month. The available granularities adjust to the selected time range, a 12-month range will not offer minute granularity.

If a dashboard number looks wrong, see Dashboard numbers look wrong.