Extract reported medication statements

Medication Adherence extracts patient-reported medication statements from a conversation transcript for pharmacist review. It organizes medication mentions, reported use, barriers, and explicit follow-up requests into an AdherenceReport.

The extraction instructions keep the report focused on what the caller actually said. Clinical decisions and follow-up actions remain with the pharmacist. The example therefore supports review of reported statements rather than deciding treatment or carrying out a requested action.

Use this example to structure a caller's account of medication use before a pharmacist reviews it. The sample transcript includes missed doses attributed to refill cost, a request for a pharmacist callback, and a second medicine mentioned without an adherence statement.

When the caller does not say whether a medicine was taken as prescribed, the extraction prompt requires an unclear status. A medication mention alone should not be treated as evidence of adherence. Use the report to preserve the caller's evidence rather than fill gaps with clinical assumptions.

The extract_report function takes a transcript string and a model name. Its structured output contains medications and follow_up_requests. Each medication record stores the caller's medication name, a mention quote, a reported status, an optional status quote, and a list of reported barriers.

Status values are as_prescribed, missed_doses, and unclear. Barrier categories cover cost, side_effect, forgetfulness, instructions, access, and other. Each barrier retains the caller's exact words, and follow-up requests are also copied exactly.

A Pydantic validator checks that the medication name is nonempty and appears within its mention quote. Any status other than unclear requires a status quote. The validator collects medication, status, barrier, and follow-up quotations and rejects empty quotations or quotations absent from the supplied transcript. The original transcript must be provided in validation context.

Instructor connects to an Anthropic provider using the model argument, while Pydantic defines and validates the report. The executable example reads ANTHROPIC_MODEL and prints indented JSON:

    report = extract_report(SAMPLE_TRANSCRIPT, os.environ["ANTHROPIC_MODEL"])

The extracted statements are intended for pharmacist review, with clinical judgment and subsequent follow-up left to that reviewer.