npx @buildinternet/releases get langfuseRecently shipped Topics trace classification, decision-model evaluator results, and skills imports from files and ZIPs.
Topics classify and summarize traces. Traces are now classified into buckets, with an adaptive map exploration view and compact workspace. All facets of a trace summarize in one call, and the summary prompt and default facets were reworked.
Decision-model evaluators expanded across results, MCP, and OpenAI. Eval decision model results were revamped, OpenAI decision models are supported, and decision-model evaluators work over MCP. Experiments annotate directly from the comparison grid, with persistent resizable charts and evaluator actions pinned while tables scroll.
Sessions render as transcripts. Session view uses trace transcripts, and session tool previews get syntax highlighting. Table rows can use a custom height. Org usage breakdown appears on every deployment, and enterprise SSO remembers the last email on the sign-in form.
Skills import from local files and ZIPs. Skills take local file and ZIP imports, unified file and folder uploads, and rendered file previews with raw source access.
AI gateway adds request correlation. Resolution context is exposed in response headers, spans, and logs, and a gateway request ID correlates requests. Full-mode inputs capture up to 5 MiB.
SDK wrap-up. The Python exporter added opt-in gzip compression (v4.17.0). The JS exporter caps oversized batches and stopped serializing each export batch twice; the LangChain integration records completion start time for streamed chat models and no longer reports LangGraph control flow as errors.
Span attribute limit checks are skipped when a span has room, and create_score request bodies are built without Fern's dict(). Plus documentation clarifying JSON-encoded metadata values in mask_otel_spans.
Adds organization-level LLM connections, records billable units in an append-only ClickHouse ledger, and shows prompt cache hit rate on traces. Fixes include keeping webhook secrets out of logs, mapping session IDs correctly for CMA agents, and rolling back Next.js and React to restore local development.
The redesigned session view is now enabled by default, and a new audit_logs table with a staging outbox relays audit events into ClickHouse. Also adds a filter and search bar for topic traces, transcript inspection from the trace view, and fixes monitors now retrying execution failures instead of disabling alerts.
LLM credentials can now be configured once at the organization level, with per-project overrides when needed.
New SDK majors are built for Langfuse v4, adding experiment links, typed metadata filters, and compressed span export. Described as a small upgrade for most users.
Password-reset emails now require Turnstile verification, and media uploads run through a pool with concurrency and byte limits. Also added per-project trace idle time in tracing_readiness_settings, Topics configured on the project's own LLM connections, media extraction before TypeScript parsing in OTel ingestion, and several native/OTel performance optimizations.
The TypeSafe System One adapter can be pointed at a custom base URL to run decision-model evaluators with SubQ Snap or Strands Decider.
The redesigned session view reads user and assistant messages across traces as one continuous conversation, with a sidebar for navigating messages, traces, and tool usage.
Spend alerts now fire for ClickHouse-billed organizations, the MCP get observations tool requires a date range, and batch score results are matched before the accepted response. Most score, session, dataset, and prompt reads were routed to the read replica.
Adds Google Cloud Storage as a blob-storage export target using a service account key or application default credentials, batch processing support for POST /scores, and skills import from local files and ZIPs with version comparison and draft previews. Also hardens ingestion with 503 Retry-After responses when S3 throttles event uploads and dedups flattened metadata paths, plus dozens of smaller web UI, sessions, charts, and RBAC fixes.
Traces can now be classified into buckets by topic, and an external media storage integration is available. Fixed a worker issue where data retention was applied without an enterprise license, and session lookups failing on NUL-byte session ids.
OpenTelemetry export batches are no longer serialized twice, reducing per-export overhead in the JS SDK.