Config Evals

Config evals are the only evaluation form you work with. Do not add, read, rewire, or reason about on-canvas evaluation nodes (EvaluationTrigger, Evaluation/checkIfEvaluating/setOutputs/setMetrics). If the user asks for those, build a config eval instead and briefly say that is how you set up evaluations.

What a Config Eval Needs

• name — a human-readable evaluation name.
• startNodeName — the node where a run begins; it is fed one test-input row. Must be a node with an incoming connection — never a trigger (see step 2).
• endNodeName — the node whose output is judged.
• dataTableId — a Data Table holding the test dataset. Create and populate it with create_data_table and insert_data_table_rows first, then link it here by id.
• metrics — one or more judged metrics (see below).

Default Procedure

1. Identify the target workflow and read it. Trace the main path from trigger to the node that produces the answer.
2. Pick the nodes:
   • startNodeName is the first node after the trigger — the node that receives the input the dataset varies. Never use the trigger itself: an eval run swaps the trigger for a dataset-driven one, so the start node must have an incoming connection or the run fails to compile. For a chat/agent workflow this is usually the agent node (often the same as endNodeName).
   • endNodeName is the node whose output you want scored (usually the AI agent or the final response node).
3. Resolve the dataset. Call list_data_tables to find an existing dataset, or create and seed one with create_data_table and insert_data_table_rows before creating the config. Never invent a dataTableId; use one returned by the data table tools.
4. Choose metrics and build the actualAnswer / expectedAnswer / userQuery expressions (see Metrics).
5. Call eval-config (action="create"), or update when changing an existing config. The tool shows an approval card automatically — call it and respect the result; do not ask for chat approval first.
6. Close with facts: evaluation name, workflow, start/end nodes, dataset name and id, and the metrics configured.

Metrics

Each metric is LLM-judged and needs a judge model: a credentialId, a model, and an outputType (numeric, the default, or boolean). Reuse an LLM credential the workflow already uses when one fits.

Do not set provider unless you know the exact chat-model node type — it is derived automatically from the credential you pass (each credential type maps to one provider). Just pick the credential and the model.

Two presets are available:

• correctness — compares the produced answer to a ground-truth answer. Requires expectedAnswer (an n8n expression resolving to the ground-truth value, typically a dataset column, e.g. ={{ $json.expected_output }}).
• helpfulness — judges the produced answer against the user’s query. Requires userQuery (an n8n expression for the input the user asked, e.g. ={{ $json.input }}).

Every metric also needs actualAnswer: an n8n expression resolving to the workflow’s produced answer at the end node, e.g. ={{ $json.output }}.

userQuery and expectedAnswer name dataset columns (the input the user asked; the ground-truth answer). actualAnswer names a field of the workflow’s produced output. Write all of them as ={{ $json.<name> }} — the evaluation reads dataset columns from the dataset row and actualAnswer from the end node automatically. Do not reference the trigger or any node by name.

Expression fields must begin with =

actualAnswer, userQuery, and expectedAnswer are n8n expressions — they read a value out of each test row at runtime. The leading = is what tells n8n to evaluate the {{ … }} template. Without it the string is stored as literal text: the field shows {{ $json.output }} verbatim and the judge scores that raw string instead of the resolved value.

• Correct: ={{ $json.output }}, ={{ $json.expected_output }}
• Wrong: {{ $json.output }} (no = → treated as fixed text)

Only add = when the value references workflow data via {{ … }}. A genuinely fixed constant (rare for these fields) is written as plain text without =.

Pick correctness when the dataset has a known right answer to compare against; pick helpfulness when there is no single ground truth and quality is judged relative to the request. Use prompt only to override the default judge prompt.

Dataset Boundary

• Build the dataset with create_data_table and insert_data_table_rows: one column for each input the evaluation varies, plus a ground-truth column when using correctness.
• The config only references the dataset by dataTableId; the eval-config tool does not create or populate rows. If no suitable dataset exists, create one first, then create the config.
• Do not weaken the evaluation to fit a thin dataset — seed the dataset to match the metrics, or ask the user for the expected answers.

More Detail

Load the config-eval-playbook reference for tool-call recipes, worked examples, and output shapes.