feat: CV Batches system with AI extraction, matching, and slide-in chat

ai_service.py

@@ -455,4 +455,161 @@ def chat(user_message: str, conversation_history: list = None, context: str = ""
     
     messages.append({"role": "user", "content": user_message})
     
-    return llm_chat(messages, temperature=0.5, max_tokens=2000)
+    return llm_chat(messages, temperature=0.5, max_tokens=2000)
+
+
+# ============================================================
+# BATCH FUNCTIONS — extract requirements, match candidates, batch chat
+# ============================================================
+
+EXTRACT_REQUIREMENTS_PROMPT = """You are a requirements analyst. Analyze the following document and extract the job positions and their requirements.
+
+Return ONLY a valid JSON object with this structure:
+{
+  "batch_name": "Short name for this batch (from the document title or first heading)",
+  "description": "Brief description of what this batch is for (1-2 sentences)",
+  "positions": [
+    {
+      "job_title": "Position title",
+      "num_positions": 1,
+      "required_skills": ["skill1", "skill2"],
+      "required_years": 5,
+      "required_certs": ["cert1"],
+      "nice_to_have": ["skill3"],
+      "disqualifiers": ["something that would disqualify a candidate"],
+      "description": "Brief description of the role"
+    }
+  ]
+}
+
+Rules:
+- Extract ALL positions mentioned in the document.
+- If a field is not specified, use empty string "" or empty array [].
+- Be precise about required skills — distinguish must-haves from nice-to-haves.
+- If no positions are found, return an empty positions array.
+- Return ONLY the JSON, no other text.
+
+Document text:
+---
+__DOC_TEXT__
+---"""
+
+def extract_requirements(doc_text: str) -> dict:
+    """Extract job requirements from a document using AI."""
+    max_chars = 15000
+    if len(doc_text) > max_chars:
+        doc_text = doc_text[:max_chars] + "\n[...truncated...]"
+    
+    prompt = EXTRACT_REQUIREMENTS_PROMPT.replace("__DOC_TEXT__", doc_text)
+    messages = [
+        {"role": "system", "content": "You are a requirements analyst that outputs only valid JSON. Make sure all string values are properly escaped."},
+        {"role": "user", "content": prompt}
+    ]
+    
+    for max_tokens in [4000, 6000, 8000]:
+        try:
+            resp = llm_chat(messages, temperature=0.1, max_tokens=max_tokens)
+            return extract_json(resp)
+        except (ValueError, json.JSONDecodeError):
+            if max_tokens == 8000:
+                raise
+            continue
+
+
+MATCH_PROMPT = """You are a CV matching specialist. You are given a job position's requirements and a list of candidates with their skills, experience, education, and certifications.
+
+For each candidate, determine:
+1. A match score (0-100) based on how well they meet the requirements
+2. Whether they are a good fit (score >= 60)
+3. A brief reasoning for the match or mismatch
+4. If they are a close match (score 50-79), suggest how their CV could be realigned (reworded) to better fit — without fabricating skills they don't have
+
+Return ONLY a valid JSON array of matching candidates:
+[
+  {
+    "candidate_id": "the UUID",
+    "candidate_name": "First Last",
+    "match_score": 85,
+    "fit": true,
+    "reasoning": "Strong match: has 7 years Python, AWS cert, Kubernetes experience",
+    "realignment_suggestion": "Emphasize microservices architecture experience over generic 'distributed systems' wording; reorder skills to put AWS and Kubernetes first"
+  }
+]
+
+Only include candidates with match_score >= 40. Sort by match_score descending.
+Return ONLY the JSON array, no other text.
+
+Position Requirements:
+__POSITION__
+
+Candidates:
+__CANDIDATES__"""
+
+def match_candidates_for_position(position: dict, candidates: list) -> list:
+    """Match candidates against a single position's requirements using AI."""
+    # Build a compact summary of each candidate for the LLM
+    cand_summaries = []
+    for c in candidates:
+        skills = ", ".join([s.get("skill_name", "") for s in c.get("skills", [])[:15]])
+        exp = "; ".join([f"{e.get('position','')} at {e.get('company','')}" for e in c.get("experience", [])[:5]])
+        certs = ", ".join([cert.get("name", "") for cert in c.get("certifications", [])[:5]])
+        edu = "; ".join([f"{e.get('degree','')} {e.get('field_of_study','')} at {e.get('institution','')}" for e in c.get("education", [])[:3]])
+        
+        cand_summaries.append(f"""Candidate ID: {c.get('id', 'N/A')}
+Name: {c.get('first_name', '')} {c.get('last_name', '')}
+Skills: {skills}
+Experience: {exp}
+Certifications: {certs}
+Education: {edu}
+Summary: {c.get('summary', '')[:200]}""")
+    
+    position_json = json.dumps(position, indent=2)
+    candidates_text = "\n\n---\n\n".join(cand_summaries)
+    
+    prompt = MATCH_PROMPT.replace("__POSITION__", position_json).replace("__CANDIDATES__", candidates_text)
+    messages = [
+        {"role": "system", "content": "You are a CV matching specialist that outputs only valid JSON. Make sure all string values are properly escaped."},
+        {"role": "user", "content": prompt}
+    ]
+    
+    for max_tokens in [4000, 6000, 8000]:
+        try:
+            resp = llm_chat(messages, temperature=0.1, max_tokens=max_tokens)
+            result = extract_json(resp)
+            if isinstance(result, list):
+                return result
+            elif isinstance(result, dict) and 'candidates' in result:
+                return result['candidates']
+            return [result]
+        except (ValueError, json.JSONDecodeError):
+            if max_tokens == 8000:
+                raise
+            continue
+    return []
+
+
+BATCH_CHAT_SYSTEM = """You are an AI assistant helping a user manage a CV batch. You have context about the batch (name, description, positions, and matched candidates).
+
+You can help the user with:
+1. Discussing which candidates are best for specific positions
+2. Suggesting CV realignment — rewording a candidate's experience to better match a position (without fabricating)
+3. Answering questions about candidate qualifications
+4. Recommending which candidates to approve or remove
+
+Be concise and direct. When suggesting changes, be specific about what to change and why.
+
+Batch context:
+__BATCH_CONTEXT__"""
+
+def batch_chat(user_message: str, batch_context: str, conversation_history: list = None) -> str:
+    """Context-aware chat for a specific batch."""
+    system_content = BATCH_CHAT_SYSTEM.replace("__BATCH_CONTEXT__", batch_context)
+    messages = [{"role": "system", "content": system_content}]
+    
+    if conversation_history:
+        for msg in conversation_history[-10:]:
+            messages.append({"role": msg["role"], "content": msg["content"]})
+    
+    messages.append({"role": "user", "content": user_message})
+    
+    return llm_chat(messages, temperature=0.5, max_tokens=3000)