Filter Search Bar
Filter traces and observations by typing. A fast query bar with operators, full-text search, wildcards, and autocomplete.
Say you are debugging an agent and you want the ERROR-level tool calls from the last hour, in production, slower than two seconds, on your checkout flow. Until now that meant clicking through a stack of sidebar dropdowns. Now you can just type it — the bar suggests fields and values as you go, so there's nothing to memorize:
level:ERROR type:TOOL environment:production latency:>2 name:*checkout*The Filter Search Bar turns one line of text into the same filters you would otherwise assemble in the sidebar, with autocomplete the whole way.
The Filter Search Bar runs on the Langfuse v4 data model. It is currently available on Langfuse Cloud only; turn on Fast on the Observations or Traces table to use it. Open-source support follows once the v4 schema migration is ready for self-hosted deployments.
What you can type
- Fields and values:
level:ERROR,environment:production,user:alice - Short aliases: type less with
env(environment),user(user id),session(session id),model(model name),cost(total cost),tokens(total tokens),tags(trace tags), plusttftandtps - Operators:
latency:>2,cost:>=0.01,startTime:>2026-06-01 - Full-text search: a bare word or phrase like
refund failedmatches across ids, names, input and output; scope it withinput:oroutput: - Wildcards:
name:*checkout*(contains),name:checkout*(starts with) - Negation:
-environment:production - Metadata and scores:
metadata.region:eu,scores.accuracy:>0.8 - Any-of for one field:
level:(ERROR OR WARNING)
Type a field name and autocomplete suggests operators and observed values; press Enter to apply. The bar runs next to the existing filter sidebar and time-range selector.
How it works
The bar parses your text into an AST and compiles it down to the same filter representation the sidebar and the filters API already use. Two consequences: your existing Saved Views keep working unchanged, and the full query is serialized into the URL, so sending someone the link reproduces the exact filtered view (a frequently requested workflow).
Free-text terms are matched with ClickHouse full-text search, so search stays fast on projects with millions of traces. A bare term searches ids, names, input and output; input: and output: scope it to one payload.
The story behind it
This one took a while: a custom grammar with a hand-written parser, a text input rebuilt from scratch on a single contenteditable, and a lot of edge cases. We wrote up the whole path, dead ends included, in a thread: how we built the Filter Search Bar.
What is next
This is the foundation for one query language across Langfuse, including nested boolean expressions and broader coverage. We would love to hear which queries you reach for most.
See the Filter Search Bar docs to learn how to use it.