Search & Query

Traceway includes a query DSL for filtering traces and spans. The same syntax works in the dashboard search bar and in API query parameters. Filters combine with AND logic — results must match every filter.

Query syntax

A query is a sequence of filters separated by spaces. Each filter uses key:value syntax, with optional comparison operators for numeric fields.

model:gpt-4o status:completed cost:>0.05

Operators

Operator	Syntax	Example	Description
Equals	`key:value`	`model:gpt-4o`	Exact string match
Quoted equals	`key:"value"`	`name:"my agent"`	Exact match with spaces
Greater than	`key:>value`	`cost:>0.10`	Numeric comparison
Less than	`key:<value`	`duration:<500`	Numeric comparison
Greater or equal	`key:>=value`	`tokens:>=1000`	Numeric comparison
Less or equal	`key:<=value`	`cost:<=0.01`	Numeric comparison
Negation	`-key:value`	`-model:gpt-4o`	Exclude matching results

Bare text (without a key: prefix) performs full-text search across span names, inputs, and outputs:

summarize document

This finds spans where the name, input, or output contains both "summarize" and "document".

Filter parameters

String filters

Key	Matches on	Example
`model`	Model name	`model:claude-3-5-sonnet`
`provider`	Provider name	`provider:anthropic`
`kind`	Span kind	`kind:llm_call`
`status`	Span status	`status:failed`
`name`	Span name (substring)	`name:generate`
`trace`	Trace ID (exact)	`trace:01234567-89ab-cdef-0123-456789abcdef`
`session`	Session ID (exact)	`session:sess_abc123`

The kind field accepts: llm_call, custom, fs_read, fs_write, tool_call, retrieval, embedding.

The status field accepts: running, completed, failed.

Numeric filters

Key	Unit	Example
`cost`	USD	`cost:>0.01`
`tokens`	Count (input + output)	`tokens:>5000`
`input_tokens`	Count	`input_tokens:>2000`
`output_tokens`	Count	`output_tokens:>1000`
`duration`	Milliseconds	`duration:>3000`

Date filters

Key	Format	Example
`since`	ISO 8601 or relative	`since:2024-06-15T00:00:00Z`
`until`	ISO 8601 or relative	`until:24h`

Relative time values: 5m (5 minutes), 1h (1 hour), 24h (24 hours), 7d (7 days), 30d (30 days).

Content filters

Key	Searches in	Example
`input`	Span input data	`input:"system prompt"`
`output`	Span output data	`output:"I don't know"`

Content filters use substring matching. Wrap multi-word values in double quotes.

Autocomplete

The dashboard search bar provides autocomplete suggestions as you type. Start typing a key name and the search bar shows available keys:

mo  →  model:
pr  →  provider:
st  →  status:
ki  →  kind:
co  →  cost:

After typing a key and colon, autocomplete shows known values from your data:

model:  →  gpt-4o, claude-3-5-sonnet, gpt-4o-mini, ...
status: →  completed, failed, running
kind:   →  llm_call, custom, tool_call, ...

Press Tab to accept a suggestion, or keep typing to narrow the list.

Example queries

Find all Claude calls that cost more than 10 cents:

model:claude-3-5-sonnet cost:>0.10

Find failed spans in the last hour:

status:failed since:1h

Find slow tool calls (over 10 seconds):

kind:tool_call duration:>10000

Find spans where the output mentions a specific error:

output:"rate limit exceeded" provider:openai

Exclude a specific model from results:

kind:llm_call -model:gpt-4o-mini

Find high-token conversations in a session:

session:sess_abc123 tokens:>5000

Saving searches

Click the bookmark icon next to the search bar to save the current query. Saved searches appear in the dropdown when you focus the search bar.

Each saved search stores:

The full query string
A display name (editable)
The page it was saved from (Traces or Spans)

Saved searches are shared across the project — any team member can see and use them. To delete a saved search, hover over it in the dropdown and click the remove icon.

On this page