Langfuse Integration
Connect Langfuse for AI application traces, prompts, feedback, latency, and usage analysis.
Connect Langfuse to give Brief context about your AI application's requests. Brief can read traces, generation observations, prompt templates, and available usage and cost measures. This helps you investigate failures and latency, understand how people use an AI feature, and review feedback alongside activity.
Connect Langfuse
- In your Langfuse project's settings, copy its Secret Key and Public Key. Use keys from the project you want Brief to read.
- Open Integrations, find Langfuse, and select Connect.
- Enter Secret Key and Public Key. Leave Host (optional) blank for
https://cloud.langfuse.com, or enterhttps://us.cloud.langfuse.comfor a project in the US Cloud region. - Select Save.
This is a shared Brief workspace connection. Brief's read clients use those two Langfuse Cloud hosts. Another supplied host falls back to the default Cloud host, so a self-hosted address will not direct Brief's reads to that instance.
Explore traces and AI activity
You can ask Brief about recent traces, observations, prompt templates, token use, costs, or latency where Langfuse provides that information. For example, ask “Show recent traces with errors” or “Which generation observations used the most tokens?” Brief's chat tools return trace and observation metadata, timing, and usage, but omit raw input and output payloads. Cost is available on some read paths; observation cost is unavailable when Brief uses Langfuse's newer observation API. Open Langfuse itself to inspect full payloads.
Open User Intelligence → Behavior. If the view is empty, select Have Brief Analyze Usage; use Sync to refresh an existing view. When data is available, open AI Chat Engagement for aggregate AI activity. Brief uses traces as sessions, generation observations as the message count, and trace user IDs for user and return-rate measures. An anonymous trace is treated as a separate user, so add user IDs to your instrumentation if you need meaningful return use. Recognized feedback in trace metadata can contribute to satisfaction. Feedback tags can also contribute when Brief can use Langfuse's older trace API; they are unavailable through its newer observation API. Generation error levels and timing contribute to error rate and average latency when present. A generation is an LLM call, so the count may differ from user-visible chat turns. Brief reads up to 500 traces and 500 generation observations for the activity card. When a result set reaches that limit, the card marks its metrics approximate and does not show trend comparisons.
Brief can also derive prompt-related research signals from traces whose prompt text is available. If the project has no accessible prompt content, those signals may be empty even when the trace count is nonzero. See Research Signals for that view.
Common problems
- Connection or reads fail: Check that both keys are correct and belong to the intended project. Check the connection card in Integrations for a health warning. Ask a workspace admin to use Reconnect if the shared connection is failing. An admin can also use Disconnect to remove it.
- Traces are missing: Confirm the keys' project and Host (optional) region, then check the time range and data in Langfuse. An unsupported host value falls back to the default Cloud host rather than showing a “host not recognized” error.
- Return rate or satisfaction looks empty: Add stable user IDs to traces and record recognized feedback markers in trace metadata if you want satisfaction. Feedback tags work only when Brief can use Langfuse's older trace API.
- Latency is unavailable or error rate is 0%: Latency needs generation observations with timing data. A 0% error rate means none of the retrieved generation observations was marked
ERROR; check your error instrumentation if you expected errors. - Some reads are unavailable for a newer Langfuse organization: Brief can use Langfuse's newer observation API for trace and observation windows when older endpoints are unavailable. Session listing and daily metrics have no replacement through that path, and prompt text may be unavailable for research signals. Inspect those details in Langfuse when Brief reports the read as unavailable.
If another AI analytics source is connected, check the source name on AI Chat Engagement to see which one supplied the metrics. See Arize for another AI observability source or all integrations for more connections.