166 lines
6.7 KiB
Python
166 lines
6.7 KiB
Python
"""Persona draft generation contract helpers."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import hashlib
|
|
from typing import Any
|
|
|
|
from .contracts.engine_gateway import GenerateResponse, structured_payload_from_response
|
|
from .persona_read_model import PersonaDraftGenerateRequest, PersonaDraftPayload
|
|
|
|
PERSONA_DRAFT_PROMPT_BUNDLE_ID = "persona-draft-rag"
|
|
PERSONA_DRAFT_PROMPT_BUNDLE_VERSION = "2026-06-28.1"
|
|
PERSONA_DRAFT_SYSTEM_PROMPT = (
|
|
"출력은 반드시 structured_schema를 따른다. code는 P숫자 형식을 선호하되 "
|
|
"힌트가 없으면 빈 문자열 대신 임시값 P로 둔다. source_provenance에는 "
|
|
"RAG source_id와 첨부 근거 기반 초안임을 남긴다. evidence chunk id를 "
|
|
"임상 필드 본문에 그대로 노출하지 않는다."
|
|
)
|
|
PERSONA_DRAFT_USER_PROMPT_PREAMBLE = (
|
|
"너는 Vignette 임상 콘텐츠 저작 보조자다. 아래 RAG 근거 청크만 바탕으로 교육용 "
|
|
"가상내담자 페르소나 초안을 만든다. 첨부 원문은 KB 문서가 SSOT이며, 근거 밖 내용을 "
|
|
"임의로 꾸며 핵심 임상 정보처럼 쓰지 않는다. 실제 개인정보는 이미 마스킹됐으며, "
|
|
"원문 표현을 복사하지 말고 "
|
|
"범주화·합성화된 임상 훈련용 설정으로 변환한다. CCD/DSM/역린은 런타임 내부 설정이므로 "
|
|
"내담자 발화에 직접 노출되지 않는 형태로 작성한다."
|
|
)
|
|
|
|
|
|
def persona_draft_prompt_bundle() -> dict[str, str]:
|
|
payload = "\n".join(
|
|
[
|
|
PERSONA_DRAFT_PROMPT_BUNDLE_ID,
|
|
PERSONA_DRAFT_PROMPT_BUNDLE_VERSION,
|
|
PERSONA_DRAFT_SYSTEM_PROMPT,
|
|
PERSONA_DRAFT_USER_PROMPT_PREAMBLE,
|
|
]
|
|
)
|
|
return {
|
|
"id": PERSONA_DRAFT_PROMPT_BUNDLE_ID,
|
|
"version": PERSONA_DRAFT_PROMPT_BUNDLE_VERSION,
|
|
"hash": hashlib.sha256(payload.encode("utf-8")).hexdigest()[:12],
|
|
}
|
|
|
|
|
|
def persona_generation_schema() -> dict[str, Any]:
|
|
return {
|
|
"type": "object",
|
|
"additionalProperties": False,
|
|
"properties": {
|
|
"draft": {
|
|
"type": "object",
|
|
"additionalProperties": False,
|
|
"properties": {
|
|
"code": {"type": "string"},
|
|
"display_name": {"type": "string"},
|
|
"difficulty": {"type": "string", "enum": ["easy", "moderate", "hard"]},
|
|
"theory_target": {"type": "array", "items": {"type": "string"}},
|
|
"demographics": {"type": "object"},
|
|
"presenting": {"type": "object"},
|
|
"history": {"type": "object"},
|
|
"big5": {"type": "object"},
|
|
"resistance": {"type": "object"},
|
|
"speech_style": {"type": "object"},
|
|
"affect_baseline": {"type": "object"},
|
|
"ccd": {"type": "object"},
|
|
"dsm5_dimensional": {"type": "object"},
|
|
"triggers": {"type": "object"},
|
|
"source_provenance": {"type": "string"},
|
|
"is_synthetic": {"type": "boolean"},
|
|
},
|
|
"required": [
|
|
"code",
|
|
"display_name",
|
|
"difficulty",
|
|
"theory_target",
|
|
"demographics",
|
|
"presenting",
|
|
"history",
|
|
"big5",
|
|
"resistance",
|
|
"speech_style",
|
|
"affect_baseline",
|
|
"ccd",
|
|
"dsm5_dimensional",
|
|
"triggers",
|
|
"source_provenance",
|
|
"is_synthetic",
|
|
],
|
|
},
|
|
"source_summary": {"type": "string"},
|
|
"warnings": {"type": "array", "items": {"type": "string"}},
|
|
},
|
|
"required": ["draft", "source_summary", "warnings"],
|
|
}
|
|
|
|
|
|
def persona_generation_payload_from_response(response: GenerateResponse) -> dict[str, Any]:
|
|
return structured_payload_from_response(response) or {}
|
|
|
|
|
|
def coerce_persona_generated_draft(
|
|
payload: dict[str, Any],
|
|
request: PersonaDraftGenerateRequest,
|
|
) -> PersonaDraftPayload:
|
|
raw = payload.get("draft") if isinstance(payload.get("draft"), dict) else payload
|
|
if not isinstance(raw, dict):
|
|
raw = {}
|
|
theory_target = raw.get("theory_target")
|
|
theory_values = (
|
|
[str(item).strip().lower() for item in theory_target if str(item).strip()]
|
|
if isinstance(theory_target, list)
|
|
else [value.strip().lower() for value in request.theory_target if value.strip()]
|
|
)
|
|
code = str(raw.get("code") or request.code_hint or "").strip().upper()
|
|
display_name = str(raw.get("display_name") or request.display_name_hint or "자료 기반 새 페르소나").strip()
|
|
difficulty = str(raw.get("difficulty") or request.difficulty)
|
|
if difficulty not in {"easy", "moderate", "hard"}:
|
|
difficulty = request.difficulty
|
|
return PersonaDraftPayload(
|
|
code=code or "P",
|
|
display_name=display_name,
|
|
difficulty=difficulty, # type: ignore[arg-type]
|
|
theory_target=theory_values or ["humanistic"],
|
|
demographics=_json_object(raw.get("demographics")),
|
|
presenting=_json_object(raw.get("presenting")),
|
|
history=_json_object(raw.get("history")),
|
|
big5=_float_dict(raw.get("big5")) or {"O": 0.5, "C": 0.5, "E": 0.5, "A": 0.5, "N": 0.5},
|
|
resistance=_float_dict(raw.get("resistance"))
|
|
or {
|
|
"base_resistance": 0.5,
|
|
"unlock_rate": 0.1,
|
|
"decay_floor": 0.05,
|
|
"silence_prob": 0.15,
|
|
"deflection_prob": 0.25,
|
|
},
|
|
speech_style=_json_object(raw.get("speech_style")),
|
|
affect_baseline=_float_dict(raw.get("affect_baseline"))
|
|
or {
|
|
"negative_affect": 0.45,
|
|
"hopelessness": 0.2,
|
|
"anhedonia": 0.2,
|
|
"sleep": 0.2,
|
|
"anxiety": 0.35,
|
|
"suicide_ideation_stage": 1,
|
|
},
|
|
ccd=_json_object(raw.get("ccd")),
|
|
dsm5_dimensional=_json_object(raw.get("dsm5_dimensional")),
|
|
triggers=_json_object(raw.get("triggers")),
|
|
source_provenance=str(raw.get("source_provenance") or f"masked {request.source_kind}"),
|
|
is_synthetic=bool(raw.get("is_synthetic", True)),
|
|
submit_for_review=False,
|
|
)
|
|
|
|
|
|
def _float_dict(value: Any) -> dict[str, float]:
|
|
if not isinstance(value, dict):
|
|
return {}
|
|
result: dict[str, float] = {}
|
|
for key, item in value.items():
|
|
if isinstance(item, (int, float)):
|
|
result[str(key)] = float(item)
|
|
return result
|
|
|
|
|
|
def _json_object(value: Any) -> dict[str, Any]:
|
|
return value if isinstance(value, dict) else {}
|