Query Explorer

Overview

Query Explorer is PublicRisk.ai's AI-powered semantic search tool that provides instant answers from your document library with source citations and context.

Features

Natural Language Queries

Ask questions in plain English - no special syntax required.

Example Queries:

✓ "What flood mitigation measures are required for Zone AE properties?"
✓ "Compare wind damage requirements between ISO and ASCE 7-22"
✓ "What's the maximum deductible for earthquake insurance in California?"
✓ "Show me all claims related to Hurricane Harvey in Houston"

DSPy ReAct Pipeline

Query Explorer uses ReAct (Reasoning + Acting) - a teacher-student optimization framework:

Benefits:

• Higher Accuracy: 92% vs 75% baseline
• Better Citations: More relevant source documents
• Faster Responses: 2-3s vs 5-8s
• Multi-turn: Remembers conversation context

Domain-Specific Adapters

Select from 35 specialized LoRA adapters trained for public sector, insurance, enterprise, and environmental domains. Each adapter is fine-tuned on 1,500-2,000 domain-specific examples.

Sample Adapters by Category:

Public Sector:

• public_education - K-12 schools and districts
• municipal_codes - Local ordinances and regulations
• law_enforcement - Police, sheriff, corrections
• emergency_management - Disaster response and preparedness
• utilities - Water, power, gas distribution

Insurance & Risk:

• insurance - General insurance operations
• liability - General and professional liability
• workers-comp - Workplace injuries and claims
• school-risk - K-12 school safety

Enterprise:

• cybersecurity - IT security, data protection
• healthcare - Medical facilities, HIPAA compliance
• financial - Credit risk, fraud prevention
• legal - Contracts, litigation
• hr-employment - Employment law, workplace safety

Environmental:

• environmental - EPA compliance, hazards
• climate - Climate change impacts
• natural_disasters - Earthquake, flood, hurricane
• nepa - National Environmental Policy Act

How to Use

// Dropdown selection
Domain: Healthcare ▼

Question: "What flood mitigation measures are required for Zone AE properties?"

Answer:
For properties in FEMA Zone AE (Special Flood Hazard Area), the following 
mitigation measures are required:

1. **Elevation Requirements** (44 CFR §60.3(c)(2))
   - Lowest floor must be at or above Base Flood Elevation (BFE)
   - Includes basements, crawl spaces, and attached garages
   
2. **Foundation Requirements**
   - Must be designed to withstand flood forces
   - Anchored to prevent flotation, collapse, or lateral movement
   
3. **Flood Vents** (if using enclosed area below BFE)
   - Minimum 1 sq. inch per 1 sq. foot of enclosed area
   - Located within 1 foot above adjacent grade

Sources:
  [1] FEMA Technical Bulletin 1-93 (page 12)
  [2] 44 CFR Part 60 - Floodplain Management
  [3] ASCE 24-14 Flood Resistant Design

Confidence: 94%

Related Questions:
  • "What's the cost of elevating a house to meet BFE?"
  • "Are there exceptions to Zone AE elevation requirements?"
  • "How do I calculate Base Flood Elevation?"

You: "What flood mitigation measures are required for Zone AE?"
AI: [Answer about elevation, foundation, vents...]

You: "What about the cost?"
AI: [Understands you mean flood mitigation costs for Zone AE]
    "Elevation costs for Zone AE properties typically range from..."

Advanced Features

Search Filters

Refine results by document type, date, or category:

// Filter panel
Document Type: [All ▼]
  □ Policies
  □ Regulations
  □ Claims
  □ Reports

Date Range: [Last Year ▼]
  ○ Last Month
  ○ Last Year
  ● All Time
  ○ Custom Range

Category: [Any ▼]
  □ Flood
  □ Earthquake
  □ Hurricane
  □ Wildfire

Bookmarks

Save important queries for later:

// Click bookmark icon on any result
[⭐ Bookmark]

// Access from sidebar
My Bookmarks (12)
  - Flood Zone AE Requirements
  - NFIP vs Private Flood Comparison
  - CA Wildfire Mitigation Standards
  ...

Export Results

Export answers in multiple formats:

• PDF - Formatted report with citations
• DOCX - Editable Word document
• Markdown - Plain text with links
• JSON - Raw data for analysis

Example PDF Export:

[Header: PublicRisk.ai Query Export]

Query: "What flood mitigation measures are required for Zone AE properties?"
Date: December 4, 2025
Adapter: General (DSPy Optimized)

Answer:
[Full AI response...]

Sources:
1. FEMA Technical Bulletin 1-93
   URL: https://www.fema.gov/...
   Excerpt: "Lowest floor must be..."
   
2. 44 CFR Part 60
   ...

Search History

View past queries and re-run them:

// History panel
Recent Queries (50)
  [Dec 4, 2025 10:15 AM] "Zone AE flood requirements"
  [Dec 4, 2025 09:42 AM] "NFIP coverage limits"
  [Dec 3, 2025 04:30 PM] "Wildfire VHFHSZ definition"
  ...

[Clear History] [Export History]

Cyber Detection & FAIR Integration (NEW)

Automatic Cyber Risk Detection

Query Explorer now automatically detects cyber security queries and suggests quantitative analysis when appropriate.

8-Tier Detection Algorithm:

1. CVE Patterns (Confidence: 95%)

   • Format: CVE-YYYY-NNNN
   • Example: "What's the risk of CVE-2024-3094?"

2. Ransomware Keywords (Confidence: 90%)

   • LockBit, BlackCat, Royal, Conti, REvil
   • Example: "How to protect against LockBit 3.0?"

3. Phishing & Social Engineering (Confidence: 85%)

   • BEC, spear phishing, credential harvesting
   • Example: "What's the cost of a phishing breach?"

4. Infrastructure Terms (Confidence: 80%)

   • SCADA, ICS, OT, critical infrastructure
   • Example: "Water utility SCADA vulnerabilities"

5. Vendor Names (Confidence: 75%)

   • Microsoft, Cisco, VMware, SolarWinds
   • Example: "Microsoft Exchange zero-day risk"

6. Threat Actor Groups (Confidence: 90%)

   • APT28, Lazarus, FIN7, DarkSide
   • Example: "FIN7 targeting financial sector"

7. Sector-Specific (Confidence: 70-85%)

   • K-12, healthcare, finance (with cyber context)
   • Example: "K-12 ransomware preparedness"

8. Financial Keywords (Confidence: 70%)

   • Cost, budget, loss, impact, insurance
   • Example: "Budget for cyber security incidents"

Smart FAIR Analysis Hints

When Query Explorer detects a cyber query with financial keywords, a smart hint appears:

┌─────────────────────────────────────────────────┐
│ 💡 Cyber Risk Detected (92% confidence)        │
│                                                 │
│ For quantitative risk analysis with Monte      │
│ Carlo simulation and ALE calculations,         │
│ try FAIR Analysis.                             │
│                                                 │
│          [📊 Run FAIR Analysis]                │
└─────────────────────────────────────────────────┘

Trigger Conditions:

• Cyber confidence ≥ 80%
• Financial keywords present: "cost", "budget", "loss", "impact", "insurance", "ALE"

What happens when you click:

1. FAIR Analysis page opens in new tab
2. Your query is pre-filled as the risk scenario
3. Cyber context passed (query type, confidence score, sector)
4. Optional: Auto-run Monte Carlo simulation

Cross-Tool Workflow Example

Scenario: User wants ransomware financial impact

Query: "What's the cost of ransomware for K-12 schools?"

✅ Cyber query detected: ransomware, confidence: 0.92
💡 Smart hint appears with "Run FAIR Analysis" button

→ Opens FAIR Analysis with pre-filled scenario
→ Shows cyber enhancement: "CISA KEV enrichment active"

Monte Carlo Simulation (10,000 runs):
- Average Annual Loss: $1,040,000
- 90th Percentile: $2,100,000
- Budget recommendation: $2.1M for insurance

→ Click "Explore in Query Explorer"
→ Ask: "What controls reduce this risk by 50%?"
→ AI provides mitigation advice with cost/benefit

Total time: 3-5 minutes for complete qualitative + quantitative analysis

FAIR Context Badges

When you navigate from FAIR Analysis back to Query Explorer, a context badge appears:

┌─────────────────────────────────────────┐
│ 📊 FAIR Analysis Context                │
│ • Scenario: Ransomware attack...        │
│ • Average Loss: $1,040,000              │
│ • 90th Percentile: $2,100,000           │
│ • Risk Range: $250K - $3.6M             │
└─────────────────────────────────────────┘

This reminds the AI of your quantitative analysis, so mitigation advice is financially calibrated.

Example:

Query: "What controls should we prioritize?"

AI Response (with FAIR context):
"Based on your ALE of $1.04M, prioritize:

1. Immutable Backups ($50K investment)
   → Reduces ALE by $340K (33% reduction)
   → ROI: 580% in year 1
   
2. EDR with Ransomware Rollback ($80K/year)
   → Reduces ALE by $425K (41% reduction)
   → ROI: 431% ongoing
   
Budget-conscious: Start with #1 for fastest payback"

Dual-Framing Analysis

Overview

Query Explorer now supports Dual-Framing based on Prospect Theory (Tversky & Kahneman, 1981) - the same query can be answered with loss-focused, gain-focused, or neutral framing to reveal how presentation affects risk perception.

Framing Theory

Tversky & Kahneman (1981) demonstrated that how you frame information dramatically changes decisions:

The Asian Disease Problem (Original Experiment):

A disease outbreak will kill 600 people. Choose between:

GAIN-FRAMED VERSION:

• Program A: 200 people will be saved (certain)
• Program B: 1/3 probability that 600 will be saved, 2/3 probability that nobody will be saved

Result: 72% chose Program A (certain gains → risk-averse)

LOSS-FRAMED VERSION:

• Program A: 400 people will die (certain)
• Program B: 1/3 probability that nobody will die, 2/3 probability that 600 will die

Result: 78% chose Program B (certain losses → risk-seeking)

Framing Modes in Query Explorer

1. Neutral Framing (Default)

Balanced, objective presentation without emotional emphasis.

Example:

Query: "What are the risks of ransomware attacks for school districts?"

Neutral Response:
"Ransomware attacks affect approximately 40% of school districts annually, 
with an average recovery cost of $1.2M per incident. Districts with incident 
response plans recover 60% faster than those without. [CISA 2024]"

When to Use:

• Initial analysis without stakeholder bias
• Technical documentation
• Regulatory compliance assessment

2. Loss-Focused Framing

Emphasizes risks, threats, failures, and negative outcomes.

Example:

Query: "What are the risks of ransomware attacks for school districts?"

Loss-Framed Response:
"⚠️ RISK-FOCUSED VIEW:
60% of school districts remain vulnerable to ransomware attacks, with 
potential losses averaging $1.2M per incident. Without proactive incident 
response plans, districts face recovery delays exceeding 3 months, resulting 
in extended operational disruptions and data exposure risks. [CISA 2024]"

When to Use:

• Justifying security investments (emphasize what you'll avoid losing)
• Crisis communication (urgency drives action)
• Regulatory compliance arguments (highlight penalties for non-compliance)

3. Gain-Focused Framing

Emphasizes benefits, opportunities, successes, and positive outcomes.

Example:

Query: "What are the risks of ransomware attacks for school districts?"

Gain-Framed Response:
"✅ OPPORTUNITY-FOCUSED VIEW:
Districts with incident response plans successfully prevent 40% of ransomware 
attacks and recover 60% faster when attacks occur. Proactive cybersecurity 
investments save an average of $1.2M per avoided incident while protecting 
student data and maintaining operational continuity. [CISA 2024]"

When to Use:

• Securing budget approval (emphasize what you'll gain/save)
• Positive stakeholder engagement (build confidence)
• Change management (focus on benefits of new approach)

UI Controls

Framing Mode Selector:

• 3-button toggle: ⚠️ Loss Frame | ⚖️ Neutral | ✅ Gain Frame
• Located above query input field
• Tooltips explain each mode's Prospect Theory foundation

Response Indicators:

• Active framing mode displayed in Alert banner
• Color-coded: Warning (orange) for loss, Info (blue) for neutral, Success (green) for gain
• Explanation of loss aversion coefficient (λ=2.25) when non-neutral

Framing Strategy Guide

Policy Communication Matrix:

Academic Foundation

Tversky, A., & Kahneman, D. (1981).
"The Framing of Decisions and the Psychology of Choice"
Science, 211(4481), 453-458.
https://doi.org/10.1126/science.7455683

Key Findings:

• Framing Effect: Same choice, different presentations → opposite preferences
• Loss Aversion: λ = 2.25 (losses hurt 2.25× more than equivalent gains)
• Risk Attitudes: Risk-averse for gains, risk-seeking for losses
• Certainty Effect: Prefer certain outcomes over probabilistically superior alternatives

Related Nobel Prize:
Daniel Kahneman received the 2002 Nobel Prize in Economics for Prospect Theory, which provides the mathematical foundation for framing effects.

Usage Guide

Practical Examples

Example 1: Cybersecurity Investment

Neutral: "40% of organizations experience at least one security incident annually, with average remediation costs of $1.2M."

Loss-Framed: "⚠️ 60% of organizations remain vulnerable to security breaches, facing potential losses averaging $1.2M per incident without proactive security investments."

Gain-Framed: "✅ Organizations with robust security programs prevent 40% of incidents, saving an average of $1.2M per avoided breach while protecting critical data."

Example 2: Flood Insurance Policy

Neutral: "NFIP residential policies cover up to $250K for structure and$100K for contents in high-risk zones."

Loss-Framed: "⚠️ Without flood insurance, homeowners risk total loss of up to $350K in structure and contents during flood events. Federal disaster assistance averages only$5K per household."

Gain-Framed: "✅ NFIP policies provide up to $350K in protection, ensuring homeowners can rebuild and replace belongings after flood events while maintaining financial security."

Best Practices

Writing Effective Queries

✅ DO:

Be Specific:

Good: "What's the maximum coverage limit for NFIP residential policies in 2025?"
Bad:  "Tell me about flood insurance"

Ask One Question at a Time:

Good: "What are the elevation requirements for Zone AE?"
      (Then follow up) "What about the cost?"
Bad:  "What are Zone AE requirements and how much does it cost and what about permits?"

Use Follow-Ups:

First:  "What is FEMA Zone AE?"
Second: "What are the insurance requirements?"
Third:  "How does this compare to Zone X?"

❌ DON'T:

Don't Use Jargon (Unless Necessary):

Bad:  "44 CFR §60.3(c)(2) applicability"
Good: "What properties must follow flood elevation rules?"

Don't Ask Multiple Questions:

Bad: "What's Zone AE and what's BFE and what about Zone X and do I need insurance?"

Choosing the Right Adapter

Quick Selection Guide:

Technical Details

RAG (Retrieval Augmented Generation)

Query Explorer uses RAG to provide accurate, cited answers:

Parameters:

• Embedding Model: nomic-embed-text (768 dim, FREE)
• Chunk Size: 1000 tokens
• Overlap: 200 tokens
• Top-K: 5 most relevant chunks
• Rerank: Yes (cross-encoder)

DSPy Optimization

Training Process:

1. Data Collection: 1000+ expert Q&A pairs
2. Labeling: Insurance professionals validate answers
3. Optimization: DSPy trains on labeled data
4. Evaluation: Test on held-out set (92% accuracy)
5. Deployment: Deploy optimized prompts to production

Metrics:

API Integration

Endpoint

https://publicrisk--dspy-optimized-service-fastapi-app.modal.run

Example Request

// POST /query
{
  "query": "What are the flood insurance requirements for commercial properties?",
  "domain_adapter": "general",
  "conversation_id": "conv-uuid-123",
  "max_sources": 5
}

Example Response

{
  "answer": "For commercial properties, flood insurance requirements vary...",
  "sources": [
    {
      "title": "FEMA Technical Bulletin 1-93",
      "url": "https://www.fema.gov/...",
      "excerpt": "Commercial buildings must...",
      "relevance": 0.94
    }
  ],
  "confidence": 0.92,
  "conversation_id": "conv-uuid-123",
  "related_questions": [
    "What's the maximum NFIP coverage for commercial properties?",
    "Are there private flood insurance options?"
  ]
}

Troubleshooting

Common Issues

Issue: "No relevant documents found"

• Cause: Query too specific or document library empty
• Solution:
  1. Broaden your query
  2. Check if documents are uploaded (Admin Portal → Documents)
  3. Try a different domain adapter

Issue: "Service timeout after 30s"

• Cause: Modal cold start (service sleeping)
• Solution: Wait 30-60s and retry. Subsequent queries will be fast (less than 3 seconds).

Issue: "Low confidence score (less than 50%)"

• Cause: Ambiguous query or insufficient document coverage
• Solution:
  1. Rephrase query to be more specific
  2. Upload relevant documents to RAG store
  3. Contact admin to train domain adapter

Issue: "Citations not clickable"

• Cause: Document not in RAG store or URL missing
• Solution: Admin must re-upload document with proper metadata

Next Steps