""" Medical Safety Layer — PDPL-compliant output filtering. Prevents definitive diagnoses, medication dosing, and adds mandatory disclaimers. """ from __future__ import annotations import re # ── Patterns that suggest definitive diagnosis ────────────────────────────── _DIAGNOSIS_PATTERNS = [ r"أنتِ?\s+مريض[ة]?\s+بـ?", r"لديكِ?\s+(مرض|سرطان|داء|متلازمة)", r"تشخيص[كِ]?\s+(هو|:)", r"مصاب[ة]?\s+بـ?\s+\w+", r"you\s+have\s+(cancer|disease|syndrome|disorder)", r"diagnosed\s+with", ] # ── Patterns that suggest specific medication dosing ───────────────────────── _MEDICATION_PATTERNS = [ r"\d+\s*(mg|ملجم|ملغ|µg|mcg|IU)\s*(يومياً|daily|مرتين|ثلاث مرات)", r"(خذ|تناول|اشرب)\s+\w+\s+\d+", r"جرعة[كِ]?\s+(هي|:)\s+\d+", r"دواء\s+\w+\s+\d+\s*(mg|ملجم)", ] # ── Emergency keywords → prepend urgent warning ────────────────────────────── _EMERGENCY_KEYWORDS_AR = [ "ألم صدر", "ضيق تنفس", "فقدان وعي", "نزيف حاد", "شلل", "تعذر التنفس", "سكتة", "نوبة قلبية", "صعوبة التنفس", ] _EMERGENCY_KEYWORDS_EN = [ "chest pain", "difficulty breathing", "loss of consciousness", "severe bleeding", "stroke", "heart attack", "paralysis", ] # ── PDPL-compliant disclaimer (وفق PDPL السعودي) ──────────────────────────── DISCLAIMER_AR = ( "\n\n⚕️ **تنبيه طبي:** هذه المعلومات لأغراض التوعية الصحية فقط " "ولا تُغني عن استشارة طبيب مختص. لا تتخذ أي قرار علاجي بناءً عليها." ) EMERGENCY_WARNING = ( "⚠️ **تحذير:** بعض هذه الأعراض قد تستدعي تقييماً طبياً عاجلاً. " "إذا كنت تعاني من أعراض حادة، توجّه فوراً لأقرب طوارئ.\n\n" ) def _contains_pattern(text: str, patterns: list[str]) -> bool: for p in patterns: if re.search(p, text, re.IGNORECASE): return True return False def _has_emergency_keywords(text: str) -> bool: t = text.lower() return any(kw in t for kw in _EMERGENCY_KEYWORDS_AR + _EMERGENCY_KEYWORDS_EN) def _soften_diagnosis(text: str) -> str: """Replace definitive diagnosis phrases with cautious language.""" replacements = [ (r"أنتِ? مريض[ة]? بـ?(\w+)", r"النتائج قد تشير إلى \1"), (r"لديكِ? (مرض|سرطان|داء)\s+(\w+)", r"النتائج قد تشير إلى \1 \2"), (r"مصاب[ة]? بـ?\s*(\w+)", r"النتائج تستدعي فحص \1"), (r"تشخيص[كِ]? هو\s+(\w+)", r"النتائج قد تشير إلى \1"), ] for pattern, replacement in replacements: text = re.sub(pattern, replacement, text, flags=re.IGNORECASE) return text def _remove_dosing(text: str) -> str: """Remove specific medication dosing instructions.""" # Replace dosing with generic advice text = re.sub( r"(خذ|تناول)\s+\w+\s+\d+\s*(mg|ملجم|ملغ)", "استشر طبيبك لتحديد الجرعة المناسبة", text, flags=re.IGNORECASE, ) text = re.sub( r"\d+\s*(mg|ملجم|ملغ)\s*(يومياً|daily|مرتين)", "(الجرعة يحددها الطبيب)", text, flags=re.IGNORECASE, ) return text def filter_analysis_report(report: dict) -> dict: """ Apply safety filtering to the full analysis report dict. Modifies: general, abnormal_details[].شرح Adds: disclaimer to general """ # Filter general assessment general = report.get("general", "") if _contains_pattern(general, _DIAGNOSIS_PATTERNS): general = _soften_diagnosis(general) if _contains_pattern(general, _MEDICATION_PATTERNS): general = _remove_dosing(general) report["general"] = general + DISCLAIMER_AR # Filter each abnormal detail explanation for item in report.get("abnormal_details", []): sharh = item.get("الشرح", "") if _contains_pattern(sharh, _DIAGNOSIS_PATTERNS): sharh = _soften_diagnosis(sharh) if _contains_pattern(sharh, _MEDICATION_PATTERNS): sharh = _remove_dosing(sharh) item["الشرح"] = sharh return report def filter_chat_response(response_text: str, user_query: str = "") -> str: """ Apply safety filtering to chat response text. - Softens definitive diagnoses - Removes medication dosing - Adds emergency warning if needed - Adds disclaimer for medical advice """ text = response_text if _contains_pattern(text, _DIAGNOSIS_PATTERNS): text = _soften_diagnosis(text) if _contains_pattern(text, _MEDICATION_PATTERNS): text = _remove_dosing(text) # Add emergency warning if user asked about emergency symptoms prefix = "" if _has_emergency_keywords(user_query) and "⚠️" not in text[:50]: prefix = EMERGENCY_WARNING # Add disclaimer only for substantive medical responses (>100 chars) if len(text) > 100 and DISCLAIMER_AR not in text: text = text + DISCLAIMER_AR return prefix + text _INJECTION_PATTERNS = [ # Prompt override attempts r"ignore\s+(previous|all|above|prior)\s+(instructions?|prompts?|context)", r"disregard\s+(all|your|previous)\s+(instructions?|rules?|guidelines?)", r"you\s+are\s+now\s+(a|an|the)\s+\w+", r"(act|pretend|behave|roleplay)\s+as\s+(if\s+you\s+(are|were)\s+)?(\w+\s*){1,4}", r"forget\s+your\s+(training|guidelines?|instructions?|rules?)", r"(system\s+)?prompt\s*(:|=|is|says?)\s*[\"']", r"jailbreak", r"do\s+anything\s+now", r"dan\s+mode", # Arabic injection patterns r"تجاهل\s+(التعليمات|السياق|الأوامر)", r"أنت\s+(الآن|الان)\s+\w+", r"تصرف\s+كـ?\w+", ] _INJECTION_RE = re.compile("|".join(_INJECTION_PATTERNS), re.IGNORECASE) # Characters/sequences to strip before inserting into prompts _PROMPT_DELIMITERS = re.compile(r"(<\|.*?\|>|###|---+|===+|\[INST\]|\[/INST\]|<>|<>)") def sanitize_query(query: str, max_len: int = 1000) -> str: """ Strip prompt-injection attempts from user input before RAG/LLM insertion. Returns cleaned query; never raises. """ q = query[:max_len] q = _PROMPT_DELIMITERS.sub(" ", q) if _INJECTION_RE.search(q): # Blank out the injection fragment but keep any preceding medical text q = _INJECTION_RE.sub("[محتوى محذوف]", q) return q.strip() def check_emergency(user_query: str) -> str | None: """ Returns an immediate emergency response string if query contains emergency keywords, else None. Only fires for very clear emergency phrases. """ critical = ["ألم صدر شديد", "صعوبة تنفس حادة", "فقدان وعي", "نزيف لا يتوقف"] q = user_query.lower() if any(kw in q for kw in critical): return ( "⚠️ هذه الأعراض تستدعي التوجه فوراً لأقرب طوارئ أو الاتصال بالإسعاف.\n" "لا تنتظر ولا تقود بنفسك. أطلب المساعدة الآن." ) return None