To space

backend/api/llm_evaluator.py

@@ -5,6 +5,7 @@ All models (Gemma, Qwen, GPT) are routed via the Hugging Face OpenAI-compatible
 """
 import os
 import json
+import re
 import requests
 import openai
 
@@ -34,6 +35,34 @@ _FALLBACK_BASE = {
     "provider": "UNKNOWN"
 }
 
+def _extract_json(text: str) -> dict:
+    """Robustly extracts JSON from a string, handling markdown or surrounding text."""
+    text = text.strip()
+    
+    # Try direct parse first
+    try:
+        return json.loads(text)
+    except json.JSONDecodeError:
+        pass
+    
+    # Try finding JSON block in markdown
+    match = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, re.DOTALL)
+    if match:
+        try:
+            return json.loads(match.group(1))
+        except json.JSONDecodeError:
+            pass
+            
+    # Final attempt: search for first { and last }
+    match = re.search(r"(\{.*\})", text, re.DOTALL)
+    if match:
+        try:
+            return json.loads(match.group(1))
+        except json.JSONDecodeError:
+            pass
+            
+    raise ValueError("Response did not contain valid JSON format.")
+
 def _get_heuristic_analysis(data: dict) -> dict:
     """Provides a rule-based assessment when LLMs are unavailable."""
     scores = data.get("scores", {})
@@ -109,12 +138,13 @@ Peak Shaving Adh.    : {scores.get('ps_adherence', 0):.3f}
 Cycle Discipline     : {scores.get('cycle_discipline', 0):.3f}
 
 == INSTRUCTIONS ==
-1. Prioritize **Energy Arbitrage (Profit)** and **SOC Readiness** in your assessment.
-2. DO NOT be overly strict about Peak Shaving violations. Minor misses (<10%) are acceptable if the agent is generating strong profit and managing battery health (Cycle Discipline) well.
-3. Return ONLY a single valid JSON object with exactly these keys:
+1. **PRIMARY FOCUS**: Energy Arbitrage (Profit/Reward Score). This is the most important metric.
+2. **SECONDARY FOCUS**: SOC Readiness. Managing battery state before peak hours is vital.
+3. **LOW IMPORTANCE**: Frequency Regulation and Peak Shaving. Give these very small weights in your assessment.
+4. Return ONLY a single valid JSON object with exactly these keys:
 {{
   "verdict": "Excellent|Good|Needs Improvement|Poor",
-  "score": <float between 0.0 and 1.0 representing overall system performance>,
+  "score": <float between 0.0 and 1.0 representing overall system performance based on the focus above>,
   "summary": "<2-3 sentence executive summary>",
   "strengths": ["<strength 1>", "<strength 2>"],
   "weaknesses": ["<weakness 1>", "<weakness 2>"],
@@ -130,7 +160,6 @@ def _get_router_analysis(data: dict, model_id: str) -> dict:
     if not api_token:
         raise ValueError("Hugging Face token not found (expected HF_TOKEN).")
     
-    # Initialize OpenAI client pointing to HF Router
     client = openai.OpenAI(
         base_url="https://router.huggingface.co/v1",
         api_key=api_token
@@ -142,28 +171,30 @@ def _get_router_analysis(data: dict, model_id: str) -> dict:
         model=model_id,
         messages=[{"role": "user", "content": prompt}],
         temperature=0.3,
+        # Note: some OS models ignore response_format, hence the robust extractor
         response_format={"type": "json_object"}
     )
     
     content = response.choices[0].message.content
-    return {**json.loads(content), "available": True, "provider": f"HF:{model_id}"}
+    try:
+        parsed = _extract_json(content)
+        return {**parsed, "available": True, "provider": f"HF:{model_id}"}
+    except Exception as e:
+        raise ValueError(f"Failed to parse LLM response: {str(e)}\nRaw Content: {content[:200]}...")
 
 def get_llm_analysis(data: dict, provider: str = "GEMINI", model_name: str = None) -> dict:
     """Main entry point for LLM analysis. Routes all providers through HF Router."""
+    model_id = "unknown"
     try:
-        # Map provider to specific model ID on HF Router
         if provider == "GEMINI":
-            # Using Gemma 4 as the replacement for Gemini
             model_id = model_name or DEFAULT_GEMMA_MODEL
         elif provider == "QWEN":
             model_id = model_name or DEFAULT_QWEN_MODEL
         elif provider == "OPENAI":
-            # Routing OpenAI requests through HF Router as requested
             model_id = model_name or DEFAULT_GPT_MODEL
         elif provider == "HF":
             model_id = model_name or DEFAULT_GEMMA_MODEL
         else:
-            # Fallback for unknown providers
             model_id = model_name or DEFAULT_GEMMA_MODEL
             
         return _get_router_analysis(data, model_id)
@@ -175,5 +206,5 @@ def get_llm_analysis(data: dict, provider: str = "GEMINI", model_name: str = Non
             "available": False,
             "summary": f"Evaluation Failed: {error_msg}",
             "error": error_msg,
-            "detailed_analysis": f"The requested {provider} analysis could not be completed using model {model_id if 'model_id' in locals() else 'unknown'}.\n\nReason: {error_msg}\n\nPlease check your Hugging Face Space secrets and token permissions."
+            "detailed_analysis": f"The requested {provider} analysis could not be completed using model {model_id}.\n\nReason: {error_msg}\n\nPlease check your Hugging Face Space secrets and token permissions."
         }

frontend/src/pages/Evaluate.jsx

@@ -108,8 +108,13 @@ export default function Evaluate() {
               <div className="gap-14">
                 <div className="chart-grid">
                   <div className="overall-hero" style={{gridRow: 'span 2'}}>
-                    <div className="overall-num">{result.scores.overall.toFixed(3)}</div>
-                    <div className="overall-label">Overall Readiness Score</div>
+                    <div className="overall-num">
+                      {llmResult && llmResult.available ? llmResult.score.toFixed(3) : result.scores.overall.toFixed(3)}
+                    </div>
+                    <div className="overall-label row gap-14">
+                      {llmResult && llmResult.available && <Sparkles size={14} className="text-accent" />}
+                      {llmResult && llmResult.available ? 'AI-Verified Readiness Score' : 'Overall Readiness Score'}
+                    </div>
                     <div className="divider" style={{width: '60%'}}/>
                     <div className="row gap-14 text-2" style={{fontSize: 12}}>
                       <div>Tasks Tested: <b>{result.task.toUpperCase()}</b></div>