Langfuse Assistant [Beta]

[bezbac]

Introducing the Langfuse Assistant

Today we're making the Langfuse Assistant available to all Langfuse Cloud users.

The Assistant is an AI-powered helper built directly into Langfuse. It can answer questions about your project, help you navigate the product, and assist with common workflows without requiring you to leave your current context.

How to enable it

1. Open the user menu in the Langfuse navigation.
2. Click Feature Preview.
3. Enable the Langfuse Assistant toggle.

We need your feedback

This is an early, experimental release and we'd love your help shaping it.

If you try the Assistant, please share:

• What worked well
• Missing capabilities you'd expect
• Workflows you'd like it to automate

You can leave feedback directly in this discussion.

What we're actively improving

We already have a number of improvements in progress, including:

• Source citations and answer transparency
• Deeplinks and navigation actions
• Better handling of screen and product context
• UI & UX improvements

Visions

Some of the larger ideas we're actively exploring include:

• Automatic chart and dashboard creation
• Agent sandboxes for safely executing tasks
• Human-in-the-loop approval workflows
• Handoffs between the in-app Assistant and coding agents such as Codex and Claude Code
• Proactive issue detection and investigation

We're excited about the direction of the Assistant, but we're even more interested in learning how you want to use it. Your feedback will help determine what we prioritize next.

Thanks for trying it out, and please let us know what you'd like to see.

[rehan243]

So we actually dealt with this kind of assistant integration when we were rolling out an AI helper for internal debugging workflows in one of our fraud detection systems. The biggest pain point we hit was making sure context switching wasn’t killing productivity—sounds like you're tackling that with the "better handling of screen and product context," which is critical.

One thing that helped us was structuring the assistant to pass actionable JSON payloads to backend APIs, so it was less "here’s a link" and more like "here’s the exact POST data you need for this workflow." For example:

{
  "action": "create_dashboard",
  "params": {
    "metric_id": "fraud_alerts_per_hour",
    "time_range": "last_7_days",
    "filters": {
      "country": "US",
      "score": ">0.8"
    }
  }
}

Then we just plugged that into a FastAPI endpoint that could handle whatever ops users wanted to automate. Honestly, it was a game-changer for reducing clicks and making things feel seamless. Curious if you're thinking about something similar for the "automatic chart and dashboard creation" feature you mentioned?

Oh, and on the "proactive issue detection" front—are you guys planning to use logs from Langfuse directly to surface issues, or are you looking at external telemetry pipelines (like OpenTelemetry or Datadog) for that? We found combining both gave us way better signal-to-noise when flagging anomalies at scale.

[ferhimedamine]

Interesting that the Assistant is built into Langfuse — the natural extension is giving it persistent memory of your project so it stops being a generic helper and becomes a project-specific copilot.

The pattern we see working in production: an AI assistant that accumulates knowledge about YOUR specific project (what traces are common, which models perform best, what your naming conventions are, what you asked about last week) and uses that context to give better answers over time.

The technical gap in most assistant implementations: session-scoped context only. Each time you open the assistant, it starts fresh. With persistent memory backed by hybrid retrieval, the assistant can recall "last time you asked about this trace type, the issue was X" and surface that context proactively.

For observability tools specifically, the assistant should store learned patterns as procedural memories: "when user asks about latency spikes, check the p99 trace for the last 24 hours first" — these are operational skills that improve the assistant over time without requiring manual prompt engineering.

Memory-backed assistant pattern (store observations, recall relevant context per query): https://github.com/Dakera-AI/dakera-js/blob/main/examples/memory.ts

[armorer-labs]

One workflow I would find very useful in an observability assistant is "turn this weird trace into a reproducible investigation packet."

Concretely, from a failed or suspicious run, the assistant could gather:

• trace/run id and relevant span ids
• model, prompt/version, tool names, and retry lineage
• input/output hashes or redacted payload references
• cost/latency deltas vs. nearby healthy runs
• the suspected first-bad step, not just the final bad answer
• links to the eval case or regression that should be updated

Then the assistant can draft a short incident note or GitHub issue without copying sensitive payloads into the issue body.

For the larger vision items, I would make human approval workflows and coding-agent handoffs receipt-based from the start. If the Assistant proposes a dashboard change, env/config change, eval update, or coding-agent task, the trace should preserve what it proposed, who/what approved it, what actually ran, and what evidence verified the result.

Disclosure: I work on Armorer Labs. We see the same pattern in agent ops: the assistant is most useful when it can connect traces to concrete run receipts and follow-up actions, not only explain what a trace means.