Prescriptions

The Prescription Engine transforms abstract diagnostic findings into concrete, actionable engineering tasks. When ModelDoctor identifies a WARNING or CRITICAL finding, it queries a built-in knowledge base and attaches a structured Prescription to the final report.

The Prescription Object

class Prescription(BaseModel):
    id: str
    finding_title: str
    dimension: str
    recommendation: str
    implementation_steps: List[str]
    estimated_gains: str
    confidence: str

Field	Description
id	Unique identifier for the prescription rule (e.g., OVR-001).
finding_title	The exact title of the finding that triggered this prescription.
dimension	The diagnostic category (e.g., Overfitting Risk).
recommendation	A high-level description of what needs to change.
implementation_steps	A list of actionable engineering tasks (e.g., "Change max_depth from None to 10").
estimated_gains	The expected impact of applying the fix (e.g., "Improves generalization by 10–15%").
confidence	Heuristic likelihood that applying this fix resolves the symptom.

How Prescriptions Are Generated

After all Doctors complete their evaluations, the PrescriptionEngine scans the Report for any finding with a WARNING or CRITICAL severity. For each match, it queries its internal knowledge base (modeldoctor.prescription.knowledge_base.PRESCRIPTIONS). If a match is found by dimension, finding title, and tags, a Prescription is instantiated and attached to the report.

Reading Prescriptions

import modeldoctor as md

report = md.diagnose(model, X_train, y_train, X_test, y_test)

recs = report.prescription.all_recommendations
if recs:
    for rec in recs:
        print(f"Prescription [{rec.id}]: {rec.description}")
        print(f"  Estimated Gain: {rec.estimated_improvement}")
        for step in rec.steps:
            print(f"  - {step}")

Knowledge Base

The knowledge base is a static dictionary located at modeldoctor.prescription.knowledge_base.PRESCRIPTIONS. It maps specific diagnostic signatures to recovery plans covering the eight built-in diagnostic dimensions.

The knowledge base is not user-facing but can be extended if you build Custom Doctors that produce custom finding titles.