Scorer Node

The Scorer Node rates input text on a rubric you describe, using the Jev decision model (TypeSafe System One). It returns a probability-weighted position on your levels — a fractional number you can branch on, loop on, or rank with.

How it works

1. The node sends the input to Jev as a single Score question over your ordered level descriptions.
2. Jev returns score (0..n_levels-1, can land between levels), confidence, and the probability of each level.
3. The score is written to state under your chosen output key, mirrored into score and variables.

Parameters

• rubric (string): The scoring question (e.g. "How ready is this lead to buy?").
• levels (array, 2-10, ordered low→high): Level descriptions. Write concrete situations: "Just browsing, no signals of intent" — not "low". Numbers-only levels collapse confidence.
• output_key (string, default "score"): Where the score is stored. Also mirrored into variables[output_key] so a Condition Node can read it by name.
• input_key (string, default "user_message"): Which state field to score. Falls back to final_answer, then user_message.

Pairing

• Condition Node: branch on the score with >=, <, or the new between operator (e.g. between 1, 2 for a mid-band).
• Loop Node (quality mode): set quality_field to your output key to loop until the score crosses the threshold. The loop's max_iterations backstop still applies.

Use Cases

1. Lead qualification: score a lead, route to sales above threshold, nurture below.
2. Quality loops: score generated output each iteration until it clears a bar.
3. Prioritization: score tickets or messages to rank a queue.

Tips

• Use as many levels as you can describe distinctly; three is fine.
• A fractional score is a position, not a percentage — rescale in your logic if you need one.
• If the decision service is unavailable, no score is written; downstream conditions fall back to their default path.