Create prompts.py

prompts.py

@@ -0,0 +1,320 @@
+"""
+Prompt templates for Gemini Flash 2.5 API interactions
+Temperature settings and structured prompts for different analysis stages
+"""
+
+CLAIM_EXTRACTION_PROMPT = """
+You are an expert technical recruiter analyzing a CV for factual claims and credibility.
+Temperature: 0.1 for precision
+
+Section Type: {section_type}
+Section Text: {section_text}
+Seniority Level: {seniority_level}
+
+Task: Extract ALL factual claims with extreme precision. Focus on:
+
+1. Work Experience Claims:
+   - Job titles, companies, dates, team sizes
+   - Specific responsibilities and technologies used
+   - Quantifiable achievements (metrics, percentages, timelines)
+   - Leadership/architectural claims
+
+2. Project Claims:
+   - Project names, descriptions, outcomes
+   - Technical stack used, deployment status
+   - Team role and contribution level
+   - Measurable results (users, performance gains, cost savings)
+
+3. Skills Claims (EXCLUDING EDUCATION):
+   - Programming languages with proficiency levels
+   - Frameworks, tools, platforms
+   - Certifications with dates
+   - Domain expertise claims
+
+4. Research/Publication Claims:
+   - Paper titles, conference venues, citations
+   - SOTA claims with specific metrics
+   - Patents, open-source contributions
+
+Output JSON format:
+{{
+  "claims": [
+    {{
+      "claim_id": "unique_id",
+      "claim_text": "exact text from CV",
+      "category": "work_experience|project|skill|research",
+      "subcategory": "specific_type",
+      "quantifiable_metrics": ["list of numbers/percentages/dates"],
+      "technologies_mentioned": ["tech1", "tech2"],
+      "time_period": {{
+        "start_date": "YYYY-MM or null",
+        "end_date": "YYYY-MM or null",
+        "duration_months": "number or null"
+      }},
+      "seniority_claim": "junior|mid|senior|lead|architect|none",
+      "verifiability_level": "high|medium|low",
+      "evidence_present": "direct|contextual|none",
+      "links_artifacts": ["URLs if any"],
+      "needs_clarification": ["specific points to verify"]
+    }}
+  ],
+  "metadata": {{
+    "total_claims": "number",
+    "buzzword_density": "0.0-1.0",
+    "specificity_score": "0.0-1.0"
+  }}
+}}
+
+IMPORTANT: 
+- Extract ONLY explicit claims, not inferences
+- Mark vague claims ("worked on cutting-edge AI") with low verifiability
+- Flag role-achievement mismatches for interview
+- Note if metrics seem unrealistic for timeframe
+- Skip education verification completely
+"""
+
+EVIDENCE_VALIDATION_PROMPT = """
+You are validating evidence for CV claims.
+Temperature: 0.2 for balanced analysis
+
+Claims to validate:
+{claims_json}
+
+Full CV text for cross-reference:
+{full_cv_text}
+
+For EACH claim, assess:
+
+1. Direct Evidence:
+   - Links to repositories, portfolios, demos (check if provided)
+   - Certificates, publications (with identifiers)
+   - Company/project websites mentioned
+
+2. Contextual Evidence:
+   - Technical depth in descriptions
+   - Specific tool versions, configurations
+   - Problem-solution narratives with details
+
+3. Cross-Section Validation:
+   - Skills mentioned MUST appear in at least one project/work
+   - Dates must be consistent across sections
+   - Technologies should align with timeframes (no React before 2013)
+
+4. Metric Sanity:
+   - Is "500% growth in 1 month" realistic?
+   - Do team sizes match achievement scope?
+   - Are research metrics within known SOTA bounds?
+
+Output JSON:
+{{
+  "validations": [
+    {{
+      "claim_id": "from_input",
+      "evidence_score": 0.0-1.0,
+      "evidence_type": "direct|contextual|cross_referenced|missing",
+      "supporting_sections": ["list of CV sections with evidence"],
+      "artifacts_found": [
+        {{
+          "type": "github|publication|certificate|website",
+          "url": "if_present",
+          "needs_verification": true/false
+        }}
+      ],
+      "cross_validation": {{
+        "skill_used_in_project": true/false,
+        "dates_consistent": true/false,
+        "tech_timeline_valid": true/false
+      }},
+      "metric_analysis": {{
+        "realistic": true/false,
+        "explanation": "why realistic or not"
+      }},
+      "triangulation_result": "verified|partial|unverified|red_flag"
+    }}
+  ],
+  "consistency_score": 0.0-1.0
+}}
+"""
+
+RED_FLAG_DETECTION_PROMPT = """
+You are detecting credibility red flags in CV claims.
+Temperature: 0.2 for pattern detection
+
+Analyzed claims with validation:
+{validated_claims_json}
+
+Seniority Level: {seniority_level}
+Role Type: {role_type}
+
+Detect these RED FLAGS:
+
+1. Role-Achievement Mismatch:
+   - "Led/Architected" in junior roles or <6 month tenure
+   - Senior achievements with entry-level titles
+   - Sole credit for large team projects
+
+2. Timeline Issues:
+   - Overlapping full-time positions
+   - Technologies used before public release
+   - Impossibly short project durations for scope
+
+3. Metric Implausibility:
+   - Extreme percentages without context (500%+ improvements)
+   - SOTA claims exceeding published benchmarks
+   - Unrealistic user numbers or scale claims
+
+4. Vagueness Patterns:
+   - High buzzword density without specifics
+   - Missing metrics on all achievements
+   - No technical depth for "expert" claims
+
+5. Over-claiming:
+   - Too many "expert" level skills (>15)
+   - All projects "successful" with no challenges
+   - Pattern of superlatives without evidence
+
+Output JSON:
+{{
+  "red_flags": [
+    {{
+      "flag_id": "unique_id",
+      "severity": "critical|high|medium|low",
+      "category": "timeline|implausible|vague|overclaim|mismatch",
+      "affected_claims": ["claim_ids"],
+      "description": "specific issue",
+      "interview_probe": "suggested question to clarify",
+      "auto_reject": false,
+      "requires_proof": true/false
+    }}
+  ],
+  "credibility_score": 0-100,
+  "seniority_adjustment": "applied adjustment based on level",
+  "risk_assessment": "low|medium|high|critical"
+}}
+"""
+
+SOTA_VERIFICATION_PROMPT = """
+You are verifying research and technical achievement claims against known benchmarks.
+Temperature: 0.1 for factual accuracy
+
+Research/Technical claims:
+{research_claims_json}
+
+Verify against known SOTA (State-of-the-Art) as of {current_date}:
+
+For each claim:
+1. Identify the benchmark/dataset/metric
+2. Check if numbers exceed published SOTA
+3. Look for required context (dataset, conditions, hardware)
+4. Assess if improvement magnitude is plausible
+
+Known SOTA baselines to reference:
+- ImageNet accuracy: ~92% (2024)
+- BERT-base F1 on SQUAD: ~93%
+- GPT-3 perplexity: varies by dataset
+- Object detection mAP: ~60-65% on COCO
+
+Output JSON:
+{{
+  "sota_validations": [
+    {{
+      "claim_id": "from_input",
+      "benchmark": "identified benchmark/dataset",
+      "claimed_metric": "number",
+      "known_sota": "published baseline",
+      "exceeds_sota": true/false,
+      "has_context": true/false,
+      "missing_details": ["dataset", "evaluation protocol", "hardware"],
+      "plausibility": "plausible|unlikely|impossible",
+      "verification_status": "needs_clarification|likely_valid|red_flag",
+      "interview_questions": ["specific technical questions"]
+    }}
+  ]
+}}
+"""
+
+REPOSITORY_ANALYSIS_PROMPT = """
+Analyze repository evidence for verification.
+Temperature: 0.1
+
+Repository URL: {repo_url}
+Repository metrics: {repo_metrics}
+Claimed contributions: {claimed_contributions}
+
+Assess:
+1. Commit density and authorship
+2. First commit vs claim date alignment  
+3. README quality and documentation depth
+4. Issues/PRs linked to claimed features
+5. Code complexity matching claimed scope
+6. Dependencies matching claimed tech stack
+
+Output credibility score and specific findings.
+"""
+
+# Scoring calibration parameters
+SCORING_CONFIG = {
+    "weights": {
+        "credibility": 0.6,
+        "consistency": 0.4
+    },
+    "seniority_thresholds": {
+        "senior": {
+            "min_evidence_score": 0.7,
+            "max_buzzword_density": 0.2,
+            "min_specificity": 0.8
+        },
+        "mid": {
+            "min_evidence_score": 0.5,
+            "max_buzzword_density": 0.3,
+            "min_specificity": 0.6
+        },
+        "junior": {
+            "min_evidence_score": 0.3,
+            "max_buzzword_density": 0.4,
+            "min_specificity": 0.4
+        },
+        "intern": {
+            "min_evidence_score": 0.2,
+            "max_buzzword_density": 0.5,
+            "min_specificity": 0.3
+        }
+    },
+    "evidence_tier_weights": {
+        "doi_arxiv": 1.0,
+        "github_active": 0.9,
+        "company_blog": 0.8,
+        "personal_blog": 0.6,
+        "no_artifact": 0.3
+    },
+    "red_flag_severity_scores": {
+        "critical": -30,
+        "high": -20,
+        "medium": -10,
+        "low": -5
+    }
+}
+
+# Bias mitigation configuration
+FAIRNESS_CONFIG = {
+    "protected_attributes": [
+        "school", "university", "college", "age", "gender", 
+        "nationality", "ethnicity", "religion", "marital status"
+    ],
+    "pii_patterns": {
+        "phone": r"\+?[\d\s\-\(\)]+",
+        "email": r"[\w\.-]+@[\w\.-]+\.\w+",
+        "address": r"\d+\s+[\w\s,]+\d{5}",
+        "ssn": r"\d{3}-\d{2}-\d{4}"
+    }
+}
+
+# Interview question templates
+INTERVIEW_TEMPLATES = {
+    "unverified_claim": "You mentioned {claim}. Can you provide more details about {specific_aspect}?",
+    "metric_clarification": "You achieved {metric}. What was the baseline and methodology?",
+    "timeline_gap": "Can you walk me through your activities between {start} and {end}?",
+    "tech_depth": "You listed {technology} expertise. Can you describe a specific challenge you solved with it?",
+    "sole_credit": "You mentioned {achievement}. Who else was involved and what was your specific contribution?",
+    "sota_claim": "Your research shows {metric} performance. How does this compare to published baselines?"
+}