Langfuse Activities

Two activities connect a workflow to Langfuse for LLM evaluation:

• langfuse.run_experiment — offline evaluation. Replays a Langfuse dataset where the task and each evaluator are themselves moco workflows.
• langfuse.create_score — online evaluation. Attaches scores to the trace of the workflow that is running right now.

Setup

Credentials come from the activity input or, when omitted, from the environment:

Input field	Environment variable
public_key	LANGFUSE_PUBLIC_KEY
secret_key	LANGFUSE_SECRET_KEY
base_url	LANGFUSE_BASE_URL

A missing value is a hard error, not a silent fallback. The worker needs the langfuse package.

These are raw credentials, not secret names

This is the only provider that takes credentials inline rather than as a *_secret_key. A key written into a workflowspec passes through workflow context and history in plaintext — prefer configuring the environment variables on the worker and omitting the fields entirely.

Scoring also relies on tracing: langfuse.create_score uses the run's trace_id as the Langfuse trace id, which is correct when the deployment exports OpenTelemetry traces to Langfuse (MOCO_OTEL_ENABLE_LANGFUSE=true).

Defaults

Activity	Timeout	Max attempts
langfuse.run_experiment	3600 s	1 — an experiment must not be replayed
langfuse.create_score	30 s	3 — safe, thanks to the deterministic score id

---

langfuse.run_experiment

Runs every item of a Langfuse dataset through a task workflow, then scores each result with one or more evaluator workflows, and records the run in Langfuse.

Each dataset item's input is handed to the task workflow as its input_data, unchanged — no schema is imposed, so the dataset must be shaped to match the workflow's inputs. Each evaluator receives {input, output, expected_output, metadata} and must return a dict, or a list of dicts, carrying name and value; a missing name defaults to the evaluator's workflowspec name.

Input

Field	Type	Required	Default	Description
dataset_name	str	yes	—	Langfuse dataset to run
task	WorkflowSpecInfo	yes	—	Workflow run for each dataset item
evaluators	list[WorkflowSpecInfo]	no	[]	Workflows that score each result
experiment_name	str	no	null	Experiment name in Langfuse
run_name	str	no	null	Name of this run
description	str	no	null	Run description
metadata	dict	no	null	Metadata attached to the run
max_concurrency	int	no	null	Dataset items processed in parallel
public_key	str	no	LANGFUSE_PUBLIC_KEY	Langfuse public key
secret_key	str	no	LANGFUSE_SECRET_KEY	Langfuse secret key
base_url	str	no	LANGFUSE_BASE_URL	Langfuse endpoint

WorkflowSpecInfo

Field	Type	Required	Default	Description
name	str	no	null	Name of a deployed workflowspec
version	str	no	latest	Version to run
content	str | list[str]	no	null	The workflowspec YAML inline, instead of a deployed name

Output

Field	Type	Description
name	str | null	Experiment name
run_name	str | null	Run name
description	str | null	Run description
experiment_id	str | null	Langfuse experiment id
dataset_run_id	str | null	Langfuse dataset run id
dataset_run_url	str | null	Link to the run in the Langfuse UI
item_results	list[object]	One per dataset item: {output, evaluations, trace_id, dataset_run_id}
run_evaluations	list[object]	Run-level evaluations, each {name, value, comment, data_type, metadata}

Example

From moco-examples/moco-agent-evaluation/src/moco-agent-evaluation.yaml:

- activity:
    type: langfuse.run_experiment
    input_data:
      dataset_name: '{{dataset_name}}'
      task:
        name: moco-agent/moco-agent-cli
      evaluators:
        - name: is_valid-wfspec
        - name: is_runnable-wfspec
        - name: llm-judge
    output_name: experiment_result

Each dataset item starts a top-level workflow

On the Temporal runtime, every task and evaluator is launched as its own top-level workflow, not as a child of the caller. A large dataset is therefore a large fan-out — set max_concurrency to bound it, and size the worker fleet accordingly.

---

langfuse.create_score

Attaches one or more scores to the currently running workflow's trace. There are no trace, observation or session fields: the target is always the caller's own trace.

Supply either the single-score shorthand (name / value / …) or a scores list — never both.

data_type must agree with the value: CATEGORICAL and TEXT need a string, NUMERIC and BOOLEAN a number (booleans are stored as 1.0 / 0.0). A mismatch is rejected at parse time.

Input

Field	Type	Required	Default	Description
scores	list[ScoreInput]	no	null	Several scores at once
name	str	no	null	Score name — single-score shorthand
value	number | str | bool	no	null	Score value — single-score shorthand
data_type	enum	no	null	NUMERIC, CATEGORICAL, BOOLEAN or TEXT
comment	str	no	null	Free-text rationale
metadata	dict	no	null	Extra metadata
config_id	str	no	null	Langfuse score config to validate against
score_id	str	no	<trace_id>-<name>	Explicit score id
public_key	str	no	LANGFUSE_PUBLIC_KEY	Langfuse public key
secret_key	str	no	LANGFUSE_SECRET_KEY	Langfuse secret key
base_url	str	no	LANGFUSE_BASE_URL	Langfuse endpoint

ScoreInput

Field	Type	Required	Default	Description
name	str	yes	—	Score name
value	number | str | bool	yes	—	Score value
data_type	enum	no	null	NUMERIC, CATEGORICAL, BOOLEAN or TEXT
comment	str	no	null	Free-text rationale
metadata	dict	no	null	Extra metadata
config_id	str	no	null	Langfuse score config
score_id	str	no	<trace_id>-<name>	Explicit score id

Output

Field	Type	Description
trace_id	str	Trace the scores were attached to
scores	list[object]	The scores written, each {score_id, name}

Example

From moco-examples/moco-agent-evaluation/src/online-score.yaml:

- activity:
    type: langfuse.create_score
    name: record_scores
    condition: "{{ len(verdict) > 0 }}"
    input_data:
      scores:
        - name: correctness
          value: "{{ float(verdict.get('correctness', 0.0)) }}"
          data_type: NUMERIC
          comment: "{{ verdict.get('rationale', '') }}"
        - name: tone
          value: "{{ verdict.get('tone', 'unknown') }}"
          data_type: CATEGORICAL
    output_name: score_result

Retries upsert rather than duplicate

The score id defaults to <trace_id>-<name>, so a retry overwrites the previous score instead of adding a second one. Override score_id only if you actually want several scores of the same name on one trace.

It needs a trace

langfuse.create_score fails when the run has no trace id. Scores also only land where they can be found if the deployment exports traces to Langfuse.