Importing from Langfuse and LangSmith
Already running Langfuse or LangSmith? Point Provantix at your project and existing traces replay into the same evidence model as live telemetry. No re-instrumentation, no data loss: a backfilled trace and a streamed event are equally chained and verifiable.
How it works
An import job reads traces and runs from your observability platform, maps each span onto the Agent Action Record model (the same mapping the SDK applies to live OTel GenAI spans), and writes them through the standard ingest pipeline: schema check, signature verification boundary, chain assembly, attribution.
Imported records are tagged with their origin connector, so you can always tell a backfilled Langfuse trace from a live SDK event in Investigate and in exports.
Langfuse
Status: available. Configure the import with:
- your Langfuse host (cloud or self-hosted),
- a project-scoped API key pair (public + secret),
- the project id to backfill.
The importer paginates through traces, respects the platform's rate limits, and is resumable: an interrupted import picks up from the last completed page rather than starting over.
LangSmith
Status: rolling out. LangSmith import uses the same pipeline with an API-key based reader. If you need it now, contact [email protected] and we'll enable it for your organization.
Tags, token detail and the platform's own cost
An import carries more than the call. For each trace or run it reads:
Cost tags. A metadata key with one of the five standard names (
department,cost_center,product,project,environment) is a tag with no setup, and so is Langfuse's own environment. Anything else becomes a tag only through a mapping you give the import job, so nothing is guessed:tagMap: { metadata: { team: 'department' }, // a metadata key to a tag tagPrefixes: { 'dept:': 'department' }, // a platform tag "dept:sales" to department = sales fields: { sessionId: 'project' }, // a field of the trace or run to a tag userMetadataKey: 'user_id', // LangSmith only: the metadata key that holds the person }The platform's own tag strings are kept as evidence either way.
Model calls (Langfuse). Each generation inside a trace is imported as a model call with its model, input and output tokens, cached and reasoning counts, and latency. Langfuse does not say which provider a model belongs to and Provantix does not guess, so pass
provider(one provider for the job) orproviderByModel; a call with no provider is not priced. Langfuse'sinputandoutputleave out what it counts under its other keys, so the parts are added back, the opposite of OpenAI's convention. Turn this off withincludeGenerations: false.Cost the platform reports. Langfuse's cost and LangSmith's
total_costare kept as reported by Langfuse or reported by LangSmith, in US dollars, beside Provantix's estimate. They never replace it. Settings, Cost tab, Source coverage shows the two together.
See What each source can carry.
After the import
- Investigate shows the imported history on the same timeline as live events.
- Candidates lists identities the importer couldn't attribute automatically; claim or merge them there.
- Coverage recalculates over the union of live + imported telemetry, so your first export already reflects history, not just go-forward data.
Tips
- Import before you cut over live traffic, so chains are contiguous from day one.
- Large backfills (millions of traces) complete faster off-peak; the importer is resumable, so partial progress is never lost.
- Keep the observability platform running after cutover if it serves debugging needs; Provantix complements it, it doesn't replace it.