# How to use a Righthand with Langfuse

Use Righthand to prepare a Langfuse trace-review packet that connects a failed output with its recorded spans and generation context.

By Righthand Team · 2026-10-12

Source: https://www.righthand.ai/blog/how-to-use-righthand-with-langfuse

## Quick answer

Use Righthand to prepare a Langfuse trace-review packet that connects a failed output with its recorded spans and generation context. Observability data is evidence, not automatic causation.

Start with a reviewable deliverable, keeping Langfuse as the accepted home for its records. This guide uses an explicitly approved snapshot or browser read; it does not assume that every provider API action is exposed by your Righthand connection.

## Evidence and access

Langfuse documents traces, observations, and sessions in its data model. See the [official Langfuse reference](https://langfuse.com/docs/observability/data-model). Provider capabilities and Righthand's available tools are separate checks.

Use the approved trace export with trace IDs, observation hierarchy, timestamps, supplied scores, and owner-defined evaluation rubrics. Exclude unnecessary prompt or customer content. Verify the required tools and owner-approved scope through [Righthand integrations](/integrations); use the supplied packet when a native read is unavailable.

## Define the objects and authority

Include these fields in the working input: trace ID, session ID, observation, model field, timestamps, status, selected input/output, evaluation label. These are the requested review columns, not a promise that every provider endpoint returns them under those exact names.

Trace payloads may contain sensitive material. Limit columns and examples to the approved investigation and avoid exporting full prompts unnecessarily.

| Responsibility | Owner |
| --- | --- |
| Prepare source scope and access | Account owner |
| Organize evidence and proposed next steps | Righthand |
| Approve consequential changes | the evaluation owner |

## Illustrative first-run brief

> Use the owner-provided Langfuse snapshot or approved read-only view. Review ten owner-selected failed task traces. Compare observable errors and output checks, then draft investigation questions without changing scores or datasets. Run once for the agreed sample, using America/Los_Angeles for the review deadline and preserving source timestamps separately. Return the draft in this conversation by Friday at 3 PM Pacific for the evaluation owner to review. Include source references and missing fields. Do not write records, spend money, change permissions, or send external messages. Stop and report an unavailable source rather than substituting another account.

## Expected result and verification

A slow generation and a bad output can coexist without proving latency caused the defect. The packet keeps observations and hypotheses separate.

Follow one trace through the relevant observations and compare its output with the accepted evaluation rule. Record the source cutoff, applied filters, and items excluded from the sample. The reviewer should be able to follow each consequential finding back to an original object. A well-written summary is not evidence that a proposed action happened. If a later write is authorized, verify its specific result separately before changing the task's status.


### Illustrative input and draft output

These synthetic entries show the requested review, not live account results or a provider field schema.

| Source item | Supplied observation | Draft conclusion |
| --- | --- | --- |
| Trace A | Rubric and score supplied | Review defined assessment |
| Trace B | Score without scale | Hold quality ranking |
| Observation C | Parent trace missing | Request complete context |

## Common failure and recovery

Short-lived instrumentation may fail to flush queued events. Flag incomplete trace coverage rather than claiming missing spans never occurred.

If a score lacks its rubric or evaluator context, preserve it as an uninterpreted value and request the missing definition before ranking traces.

## FAQ

### What should the reviewer verify first?

Verify trace and observation IDs, evaluator, rubric, and source version. A numerical score without a defined scale does not establish quality or success.

### How should a repeat run be scoped?

Retain rubric and evaluator versions across reviews. Separate traces from different scoring schemes and keep redaction rules fixed when preparing another evidence packet.

For connection and permission details, use [Righthand Docs](https://docs.righthand.ai). Review [Pricing](/pricing) alongside any separate provider charges before expanding the workflow.