Update vips_classifier.py

vips_classifier.py

@@ -1,9 +1,5 @@
 """
 VoiceNote AI - VIPS Classifier
-Classifies patient information into VIPS categories using three prompt strategies.
-
-Input: English-translated anonymized text (via DeepL)
-Output: VIPS dict for each of zero_shot / few_shot / chain_of_thought
 """
 
 import logging
@@ -11,150 +7,116 @@ from config import Config
 
 logger = logging.getLogger(__name__)
 
+_VIPS_DEFINITIONS = """VIPS categories — extract ANY mention of:
+V (Wellbeing): pain levels, fatigue, nausea, dizziness, sleep quality, mood, anxiety, appetite, physical symptoms
+I (Integrity): living situation, mobility needs, personal habits, social support, preferences, independence level
+P (Prevention): mobilization plans, exercises, lifestyle factors (smoking, diet), follow-up plans, physiotherapy
+S (Safety): fall risk, allergies, medications, postoperative risks, clot risk, infection risk, safety equipment needed"""
 
-# ══════════════════════════════════════════════════════════
-# SYSTEM INSTRUCTION (shared base for all prompts)
-# ══════════════════════════════════════════════════════════
-
-_SYSTEM = """You are a clinical documentation assistant for Swedish nursing.
-Classify patient-nurse conversations into the four VIPS categories below.
-
-VIPS categories:
-- V (Wellbeing / Välbefinnande): Subjective symptoms — pain, fatigue, nausea, mood, sleep, appetite
-- I (Integrity / Integritet): Autonomy, personal habits, living situation, social relations, preferences
-- P (Prevention / Prevention): Risk factors, preventive measures, lifestyle (smoking, diet, exercise)
-- S (Safety / Säkerhet): Allergies, fall risk, medication, infection risk, safety concerns
+_RULES = """CRITICAL RULES:
+1. Read EVERY sentence carefully — extract ALL clinical details mentioned.
+2. Write "Ingen relevant information." ONLY if the category has ZERO mentions.
+3. Be specific (e.g. "Pain 3/10 at rest, 6/10 on movement" not just "pain reported").
+4. Never invent information not stated in the conversation."""
 
-Rules:
-1. ONLY include information explicitly stated in the conversation.
-2. NEVER invent or assume information that was not mentioned.
-3. If a category has no relevant information, write exactly: Ingen relevant information.
-4. Keep each category concise and clinical."""
 
+def build_prompt_zero_shot(english_text: str) -> str:
+    return f"""You are a Swedish clinical documentation specialist generating nursing notes in VIPS format.
 
-# ══════════════════════════════════════════════════════════
-# 1. ZERO-SHOT
-# ══════════════════════════════════════════════════════════
+{_VIPS_DEFINITIONS}
 
-def build_prompt_zero_shot(english_text: str) -> str:
-    """
-    Zero-shot: no examples — pure instruction only.
-    Reference: Sivarajkumar et al. (2022), HealthPrompt
-    """
-    return f"""{_SYSTEM}
+{_RULES}
 
-Patient-nurse conversation:
+Patient-nurse conversation to document:
 \"\"\"{english_text}\"\"\"
 
-Generate the VIPS nursing note:
-V:
-I:
-P:
-S:"""
+Extract all relevant clinical information and write the VIPS note now.
+Use this exact format — one line per category:
 
+V (Välbefinnande):
+I (Integritet):
+P (Prevention):
+S (Säkerhet):"""
 
-# ══════════════════════════════════════════════════════════
-# 2. FEW-SHOT
-# ══════════════════════════════════════════════════════════
 
 def build_prompt_few_shot(english_text: str) -> str:
-    """
-    Few-shot: 3 concrete examples before the task.
-    Reference: Brown et al. (2020), Language Models are Few-Shot Learners
-    """
-    return f"""{_SYSTEM}
+    return f"""You are a Swedish clinical documentation specialist generating nursing notes in VIPS format.
+
+{_VIPS_DEFINITIONS}
+
+{_RULES}
 
 --- EXAMPLE 1 ---
 Conversation: "I have a headache and feel very tired. I haven't slept well in three days."
-V: Headache reported. Severe fatigue. Sleep disturbance for three days.
-I: Ingen relevant information.
-P: Ingen relevant information.
-S: Ingen relevant information.
+V (Välbefinnande): Headache reported. Severe fatigue. Sleep disturbance for three days.
+I (Integritet): Ingen relevant information.
+P (Prevention): Ingen relevant information.
+S (Säkerhet): Ingen relevant information.
 
 --- EXAMPLE 2 ---
-Conversation: "I take Metoprolol every day. I smoke about ten cigarettes a day and drink alcohol on weekends. I'm allergic to penicillin."
-V: Ingen relevant information.
-I: Smoker (approx. 10 cigarettes/day). Alcohol consumption on weekends.
-P: Smoking cessation may be discussed. Lifestyle risk factors identified.
-S: Known allergy to penicillin. Daily medication: Metoprolol.
+Conversation: "I take Metoprolol every day. I smoke about ten cigarettes a day. I'm allergic to penicillin. I feel dizzy when I stand up."
+V (Välbefinnande): Dizziness on standing reported.
+I (Integritet): Active smoker (approx. 10 cigarettes/day).
+P (Prevention): Smoking cessation recommended. Daily medication management ongoing.
+S (Säkerhet): Known allergy to penicillin. Daily medication: Metoprolol. Fall risk due to dizziness on standing.
 
 --- EXAMPLE 3 ---
-Conversation: "I have chest pain and difficulty breathing. I feel dizzy when I stand up. I live alone and have no one to help me."
-V: Chest pain and dyspnea reported. Dizziness on standing.
-I: Lives alone. No social support available.
-P: Ingen relevant information.
-S: Acute symptoms — immediate clinical assessment required. Elevated fall risk due to dizziness.
+Conversation: "I have chest pain and difficulty breathing. I live alone. The doctor said I need to start walking tomorrow."
+V (Välbefinnande): Chest pain and dyspnea reported.
+I (Integritet): Lives alone. No social support available at home.
+P (Prevention): Mobilization plan initiated — walking planned from tomorrow per physician order.
+S (Säkerhet): Acute cardiopulmonary symptoms require assessment. Elevated fall risk. Lives alone increases safety concern.
 
---- NOW YOUR TURN ---
+--- YOUR TURN ---
 Conversation: \"\"\"{english_text}\"\"\"
 
-V:
-I:
-P:
-S:"""
-
+V (Välbefinnande):
+I (Integritet):
+P (Prevention):
+S (Säkerhet):"""
 
-# ══════════════════════════════════════════════════════════
-# 3. CHAIN-OF-THOUGHT
-# ══════════════════════════════════════════════════════════
 
 def build_prompt_chain_of_thought(english_text: str) -> str:
-    """
-    Chain-of-Thought: explicit step-by-step reasoning before output.
-    Reference: Wei et al. (2022), Chain-of-Thought Prompting Elicits Reasoning
-    """
-    return f"""{_SYSTEM}
+    return f"""You are a Swedish clinical documentation specialist generating nursing notes in VIPS format.
 
-Use this step-by-step process:
+{_VIPS_DEFINITIONS}
 
-STEP 1 — List every clinical detail mentioned in the conversation.
-STEP 2 — Assign each detail to the correct VIPS category.
-STEP 3 — Verify: Does every item come directly from the conversation? Remove hallucinated content.
-STEP 4 — Write the final VIPS note.
+{_RULES}
 
 Patient-nurse conversation:
 \"\"\"{english_text}\"\"\"
 
-STEP 1 — Clinical details identified:
-(list all relevant information)
+Follow these steps carefully:
 
-STEP 2 — VIPS mapping:
-(assign each item to V / I / P / S)
+STEP 1 — List EVERY clinical detail from the conversation (pain, symptoms, medications, plans, risks, living situation, etc.):
+-
 
-STEP 3 — Verification:
-(confirm nothing was invented)
+STEP 2 — Assign each detail to the correct VIPS category:
+V (Wellbeing) items:
+I (Integrity) items:
+P (Prevention) items:
+S (Safety) items:
 
-STEP 4 — Final VIPS Note:
-V:
-I:
-P:
-S:"""
+STEP 3 — Check: Is every item above directly from the conversation? Remove anything invented.
 
+STEP 4 — Write the final VIPS note:
+V (Välbefinnande):
+I (Integritet):
+P (Prevention):
+S (Säkerhet):"""
 
-# ══════════════════════════════════════════════════════════
-# PARSER
-# ══════════════════════════════════════════════════════════
 
 def parse_vips_response(response: str) -> dict:
-    """
-    Parse raw LLM response into a VIPS dict.
-    Handles both 'V:' and 'V (Wellbeing):' label formats.
-
-    Returns:
-        {"V": "...", "I": "...", "P": "...", "S": "..."}
-    """
     default = "Ingen relevant information."
     vips = {"V": default, "I": default, "P": default, "S": default}
 
-    # Only parse the STEP 4 section if Chain-of-Thought
     if "STEP 4" in response:
         response = response.split("STEP 4")[-1]
 
     for line in response.strip().splitlines():
         line = line.strip()
         for key in ["V", "I", "P", "S"]:
-            if line.startswith(f"{key}:") or line.startswith(f"{key} ("):
-                # Extract content after the first colon
+            if line.startswith(f"{key} (") or line.startswith(f"{key}:"):
                 content = line.split(":", 1)[-1].strip()
                 if content:
                     vips[key] = content
@@ -163,25 +125,18 @@ def parse_vips_response(response: str) -> dict:
     return vips
 
 
-# ═══════════════════════════════���══════════════════════════
-# MAIN: run all 3 strategies
-# ══════════════════════════════════════════════════════════
+def format_vips_for_display(vips: dict) -> str:
+    labels = {
+        "V": "V (Välbefinnande)",
+        "I": "I (Integritet)",
+        "P": "P (Prevention)",
+        "S": "S (Säkerhet)",
+    }
+    return "\n".join(f"{labels[k]}: {vips.get(k, 'Ingen relevant information.')}"
+                     for k in ["V", "I", "P", "S"])
+
 
 def classify_all(english_text: str, mistral_client) -> dict:
-    """
-    Run all three prompt strategies against the same English text.
-
-    Args:
-        english_text: DeepL-translated, GDPR-anonymized text
-        mistral_client: MistralClient instance
-
-    Returns:
-        {
-          "zero_shot":       {"V":..., "I":..., "P":..., "S":...},
-          "few_shot":        {"V":..., "I":..., "P":..., "S":...},
-          "chain_of_thought":{"V":..., "I":..., "P":..., "S":...},
-        }
-    """
     strategies = {
         "zero_shot":        (build_prompt_zero_shot,        Config.LLM_MAX_TOKENS_ZERO_SHOT),
         "few_shot":         (build_prompt_few_shot,         Config.LLM_MAX_TOKENS_FEW_SHOT),