Refactor: Robust JSON extraction and grounded PM commentary for Bytez AI

ai/ai_reporter.py

@@ -26,7 +26,7 @@ class AIReporter:
         self.model_path = "meta-llama/Meta-Llama-3-8B-Instruct"
         self.endpoint = f"{self.base_url}/{self.model_path}"
 
-    def _call_bytez(self, messages: list, max_tokens: int = 500, temperature: float = 0.7) -> str:
+    def _call_bytez(self, messages: list, max_tokens: int = 500, temperature: float = 0.7, top_p: float = 0.9) -> str:
         """
         Helper to make the POST request to Bytez.
         """
@@ -41,7 +41,9 @@ class AIReporter:
         payload = {
             "messages": messages,
             "max_tokens": max_tokens,
-            "temperature": temperature
+            "temperature": temperature,
+            "top_p": top_p,
+            "response_format": {"type": "json_object"} if temperature < 0.2 else None
         }
 
         try:
@@ -64,25 +66,40 @@ class AIReporter:
     def parse_intent(self, user_prompt: str) -> list:
         """
         Uses Bytez AI to map user prompt to a list of exact GICS sectors to exclude.
+        Highly deterministic (temp=0.1) with robust regex extraction.
         """
         logger.info(f"Parsing intent with Bytez for prompt: {user_prompt[:50]}...")
         
         messages = [
             {"role": "system", "content": INTENT_PARSER_SYSTEM_PROMPT},
-            {"role": "user", "content": f"Parse this prompt for sector exclusions: '{user_prompt}'"}
+            {"role": "user", "content": f"Parse: '{user_prompt}'"}
         ]
         
         try:
-            content = self._call_bytez(messages, max_tokens=100, temperature=0.0)
+            # Intent Parser uses low temperature and top_p for determinism
+            content = self._call_bytez(
+                messages, 
+                max_tokens=100, 
+                temperature=0.1
+            )
+            
             if not content:
-                logger.warning("Empty response from Bytez for Intent Parsing. Returning empty list.")
+                logger.warning("Empty response from Bytez for Intent Parsing. Defaulting to [].")
                 return []
 
-            # Clean content for JSON extraction
+            # Robust Regex Fallback Extraction
             import re
+            # Look for the JSON list [ ... ]
             match = re.search(r'\[.*\]', content.strip(), re.DOTALL)
             if match:
-                return json.loads(match.group(0))
+                extracted_json = match.group(0)
+                try:
+                    return json.loads(extracted_json)
+                except json.JSONDecodeError as je:
+                    logger.error(f"JSON Decode Error after extraction: {je}")
+                    return []
+            
+            logger.warning(f"No JSON list found in response: {content[:100]}...")
             return []
             
         except Exception as e:
@@ -100,20 +117,26 @@ class AIReporter:
         from datetime import datetime
         current_date = datetime.now().strftime("%B %d, %Y")
         
+        # Grounding check: If Tracking Error is 0, we are in replication mode
+        is_replication = attribution_report.total_active_return == 0
+        
         # Format the user prompt
         user_prompt = f"""
 Current Date: {current_date}
-INSTRUCTION: Start your commentary exactly with the header: "Market Commentary - {current_date}"
-""" + ATTRIBUTION_PROMPT_TEMPLATE.format(
-            excluded_sector=excluded_sector,
-            total_active_return=attribution_report.total_active_return * 100,
-            allocation_effect=attribution_report.allocation_effect * 100,
-            selection_effect=attribution_report.selection_effect * 100,
-            top_contributors=json.dumps(attribution_report.top_contributors, indent=2),
-            top_detractors=json.dumps(attribution_report.top_detractors, indent=2),
-            sector_positioning=json.dumps(attribution_report.sector_exposure, indent=2),
-            current_date=current_date
-        )
+Portfolio Metadata:
+- Sector Exclusions: {excluded_sector}
+- Alpha (Active Return): {attribution_report.total_active_return * 100:.2f}%
+- Total Tracking Error: {attribution_report.total_active_return * 100:.4f}%
+- Full Replication Mode: {is_replication}
+
+## DATA TABLES:
+**Contributors/Detractors**:
+{json.dumps(attribution_report.top_contributors[:5], indent=2)}
+{json.dumps(attribution_report.top_detractors[:5], indent=2)}
+
+**Sector Positioning**:
+{json.dumps(attribution_report.sector_exposure, indent=2)}
+"""
         
         if not self.api_key:
              return f"AI Commentary Unavailable. (Missing BYTEZ_API_KEY). Current Date: {current_date}"
@@ -129,7 +152,7 @@ INSTRUCTION: Start your commentary exactly with the header: "Market Commentary -
             if not commentary:
                 return "AI Commentary generation timed out or failed. Please try again."
                 
-            return commentary
+            return commentary.strip()
             
         except Exception as e:
             logger.error(f"Failed to generate Bytez report: {e}")

ai/prompts.py

@@ -18,24 +18,33 @@ GICS Sectors:
 
 ## RULES:
 1. Return ONLY a valid JSON list of strings from the 11 GICS sectors above.
-2. If the user mentions "tech", map it to "Information Technology".
-3. If the user mentions "banks" or "finance", map it to "Financials".
-4. If the user mentions "healthcare" or "pharma", map it to "Health Care".
-5. If the user doesn't want to exclude any sectors, return [].
-6. Do NOT include any explanations or extra text.
+2. OUTPUT ONLY VALID JSON. NO MARKDOWN BACKTICKS (```json). NO EXPLANATIONS.
+3. If the user mentions "tech", map it to "Information Technology".
+4. If the user mentions "banks" or "finance", map it to "Financials".
+5. If the user mentions "healthcare" or "pharma", map it to "Health Care".
+6. If the user doesn't want to exclude any sectors, return [].
 
 Example:
 User: "no tech or banks"
 Output: ["Information Technology", "Financials"]
 """
 
-SYSTEM_PROMPT = """You are a Senior Portfolio Manager at a top-tier Asset Management firm (e.g., Goldman Sachs, BlackRock). 
-Your goal is to write a concise, professional, and insightful performance commentary for a High Net Worth Application.
+SYSTEM_PROMPT = """You are a Senior Portfolio Manager at a top-tier Asset Management firm.
+You are analyzing a direct indexing portfolio. 
+
+## GROUND RULES:
+1. **The Tracking Error Rule**: If Tracking Error is 0.00%, it means we are perfectly tracking the benchmark. Do NOT invent active returns or alpha. State that the portfolio matches the benchmark exactly.
+2. **The Exclusion Rule**: If a stock or sector has "Status": "Excluded", NEVER refer to it as a "Holding". We don't own it. 
+3. **The GICS Rule**: Adhere strictly to the "Sector" field provided in the input JSON. Do not hallucinate sectors.
+4. **Data Grounding**: Do not cite any data not present in the provided JSON "Truth Tables". Rely ONLY on the provided allocation dictionary.
+
 Your tone should be:
 1. Professional and reassuring.
-2. Mathematically precise (cite the numbers).
+2. Mathematically precise (cite the numbers from the JSON).
 3. Explanatory (explain 'why' something happened).
+"""
 
+GOLDMAN_RULES = """
 ## GOLDMAN RULES (STRICT COMPLIANCE)
 1. **The Exclusion Rule**: If a stock or sector has "Status": "Excluded", NEVER refer to it as a "Holding". We don't own it. Its negative contribution is a "Missed Opportunity" or "Drag from Benchbark Rally".
 2. **The Active Return Rule**: Only call a stock a "Contributor" if its "Active_Contribution" is POSITIVE.