Truth Knowledge Base

The Truth Knowledge Base is a repository of verified facts and information that TruthVouch uses to detect hallucinations and verify AI-generated content.

What are Truth Nuggets?

Truth Nuggets are bite-sized pieces of verified information:

Characteristics:

• Factual — Verified as accurate
• Specific — Focused on a single claim
• Sourceable — Tied to authoritative sources
• Certified — Cryptographically signed
• Versioned — Track changes over time

Examples:

"The Great Wall of China is approximately 13,171 miles long."
Source: UNESCO World Heritage
Verified by: history@company.com
Certified: 2024-01-15

"Our product minimum order is 100 units."
Source: Sales Team
Verified by: sales-manager@company.com
Certified: 2024-01-01

"Anthropic was founded in 2021."
Source: Public Records
Verified by: research@company.com
Certified: 2023-12-01

Why They Matter

1. Detect Hallucinations

Compare AI responses against your verified knowledge:

AI Response:   "The Great Wall is 5,000 miles long"
Truth Nugget:  "13,171 miles long"
Result:        HALLUCINATION DETECTED ❌

2. Maintain Accuracy

Ensure AI outputs are consistent with company facts:

AI Response:   "Minimum order is 50 units"
Truth Nugget:  "Minimum order is 100 units"
Result:        ACCURACY CHECK FAILED ❌

3. Standardize Information

Prevent contradictory information in customer-facing AI:

Chatbot 1: "We're founded in 2020"
Chatbot 2: "We're founded in 2021"
Truth Nugget: "Founded in 2021"
Result: Chatbot 1 marked unreliable

Knowledge Base Structure

Categories

Organize nuggets by domain:

• Company Facts — Founded date, locations, products
• Product Information — Features, pricing, compatibility
• Compliance — Legal requirements, regulations
• Industry Data — Market size, competitor info
• Historical Facts — Verified past events

Metadata

Each nugget includes:

• Content — The fact itself
• Category — Domain classification
• Source — Where fact came from
• Verified by — Person who confirmed accuracy
• Verification date — When confirmed
• Expiration (optional) — When fact becomes outdated
• Confidence (1-5) — How certain the fact is
• Related nuggets — Linked facts

Creating Truth Nuggets

Via Dashboard

1. Navigate to Knowledge Base → Platform → Knowledge Base → Truth Nuggets

2. Click “Create Nugget”

3. Fill in details:

• Fact (required)
• Category (required)
• Source (required)
• Confidence level (1-5 stars)
• Expiration date (optional)

4. Add sources

• URL links
• Document references
• Contact information

5. Verify

• Add internal approver
• Request review from team

6. Publish

• Nugget becomes active
• Certificate generated
• Available for scanning

Via API

curl -X POST https://api.truthvouch.io/v1/nuggets \
  -H "Authorization: Bearer token_abc123..." \
  -H "Content-Type: application/json" \
  -d '{
    "content": "The Great Wall of China is approximately 13,171 miles long",
    "category": "historical_facts",
    "source": "UNESCO World Heritage",
    "confidence": 5,
    "verified_by": "history@company.com",
    "sources": [
      {
        "type": "url",
        "url": "https://whc.unesco.org/..."
      }
    ]
  }'

Bulk Import

Upload CSV with nuggets:

content,category,source,confidence,verified_by
"The Great Wall of China is 13,171 miles long",historical_facts,UNESCO,5,history@company.com
"Our minimum order is 100 units",product_info,Sales Team,5,sales@company.com
"Founded in San Francisco in 2021",company_facts,Public Records,5,hr@company.com

Dashboard:

1. Knowledge Base → Bulk Import
2. Download template
3. Fill in facts
4. Upload CSV
5. Review and approve

Managing Your Knowledge Base

View Nuggets

Dashboard:

1. Platform → Knowledge Base → Truth Nuggets
2. Filter by category, status, age
3. Search by content or source
4. Sort by confidence, date

Columns:

• Content preview
• Category
• Confidence level
• Verified by
• Verification date
• Status (active/expired/pending)
• Usage (# of scans that checked this)

Edit Nuggets

Update information:

1. Click nugget
2. Edit fields
3. Save (creates new version)
4. Old version retained (audit trail)

Deprecate:

1. Mark as “Deprecated”
2. Set expiration date
3. Scanning still uses it until expiration
4. Generates new certificate

Archive Nuggets

Move outdated nuggets to archive:

1. Select nuggets
2. Click “Archive”
3. No longer used in scanning
4. Can be restored later

Verification & Certification

Verification Workflow

1. Create Nugget State: DRAFT

2. Request Review State: PENDING_REVIEW Assigned to: Approver

3. Approver Reviews Options:

• Approve → State: VERIFIED
• Request changes → State: NEEDS_REVISION
• Reject → State: REJECTED

4. Publish State: ACTIVE Certificate generated

Cryptographic Certification

Each verified nugget receives a certificate:

{
  "certificate_id": "cert-abc123",
  "nugget_id": "nugget-xyz789",
  "content_hash": "sha256:abcdef...",
  "verified_at": "2024-01-15T10:00:00Z",
  "verified_by": "history@company.com",
  "signature": "...",
  "expires_at": "2025-01-15T10:00:00Z"
}

Signature covers:

• Content (prevents tampering)
• Verification date
• Verifier identity
• Expiration date

Validation:

• Verify signature using issuer’s public key
• Check expiration date
• Confirm content hash matches

Blockchain Transparency (Optional)

For maximum transparency, certificates can be:

1. Hashed and stored on blockchain
2. Publicly verifiable via blockchain explorer
3. Permanent audit trail

Enable in settings: Organization → Settings → Knowledge Base → Blockchain transparency

Using Nuggets for Scanning

Automatic Hallucination Detection

When Gateway scans responses:

1. Extract claims from response
2. Search knowledge base for related nuggets
3. Compare claims against nuggets
4. Calculate similarity score
5. If contradiction found:
   - Log detection
   - Create alert
   - Mark response as potentially false
6. Return risk score to application

Manual Verification

Users can manually check responses:

Dashboard:

1. Alerts → View response
2. Click “Check against Knowledge Base”
3. See matching nuggets
4. Compare with response claims
5. Mark as verified or false

API-Based Verification

curl -X POST https://api.truthvouch.io/v1/verify \
  -H "Authorization: Bearer token_abc123..." \
  -H "Content-Type: application/json" \
  -d '{
    "content": "AI response text to verify",
    "context": "conversation context"
  }'

# Response
{
  "score": 0.85,
  "matches": [
    {
      "nugget_id": "nugget-xyz789",
      "confidence": 0.92,
      "content": "The Great Wall is 13,171 miles long",
      "status": "contradicts"
    }
  ],
  "hallucination_risk": 0.8
}

Analytics & Insights

Knowledge Base Metrics

Dashboard: Platform → Analytics → Knowledge Base

Metrics:

• Total nuggets
• Nuggets by category
• Verification coverage
• Most-used nuggets
• Detected contradictions

Scanning Impact

See how nuggets affect scanning:

• % of responses checked against nuggets
• Hallucinations detected per nugget
• False positive rate
• Time to detect hallucination

Usage Reports

Generate reports:

1. Analytics → Reports
2. Select “Knowledge Base Usage”
3. Choose date range
4. View:
   • Nuggets verified
   • Contradictions found
   • Response modifications
   • Team contributions

Best Practices

1. Keep Nuggets Fresh

• Review quarterly
• Update dates when facts change
• Archive outdated information
• Monitor expiration dates

2. Be Specific

Good: “Apple Inc. was founded on April 1, 1976” Bad: “Apple is an old company”

3. Cite Sources

• Link to authoritative sources
• Include verification details
• Track who verified
• Note verification date

4. Organize by Category

• Create category hierarchy
• Use consistent naming
• Train team on structure
• Update categories as needed

5. Get Team Input

• Invite subject matter experts
• Use approvers from relevant teams
• Track verification history
• Encourage collaboration

Troubleshooting

Nugget Not Detecting Hallucinations

Possible causes:

1. Response doesn’t directly contradict nugget
2. Similarity threshold too high
3. Nugget is too vague
4. Response is mostly correct with minor error

Solution:

• Review response vs nugget
• Adjust confidence/specificity
• Split vague nuggets into specific ones
• Adjust similarity threshold

High False Positive Rate

Possible causes:

1. Nuggets are too broad
2. Confidence scores too high
3. Contradictions are acceptable
4. Different phrasings are marked as contradictions

Solution:

• Lower confidence for ambiguous nuggets
• Review and adjust contradicting nuggets
• Mark acceptable variations
• Tune similarity algorithm

Nuggets Not Used in Scanning

Check:

1. Is knowledge base enabled? (Settings → Scanning)
2. Are nuggets verified/active?
3. Is response actually checked? (May need broader category)
4. Is similarity high enough? (Adjust threshold)

Integration Examples

Company with Multi-Product Lines

Create nuggets for each product:

Category: Product Info
- Product A: Min order 100 units
- Product B: Min order 50 units
- Product C: Min order 200 units

AI chatbot scans responses against product-specific nuggets.

Compliance-Heavy Organization

Create compliance nuggets:

Category: GDPR Compliance
- Max data retention: 90 days
- Required consent: Must be explicit
- Data processors: Must sign DPA

Category: HIPAA Compliance
- Minimum encryption: AES-256
- Access logs: Required for all systems

Gateway compares AI responses about compliance.

Global Organization

Create region-specific nuggets:

Category: Company Facts (US)
- Founded: 1995 in San Francisco

Category: Company Facts (EU)
- Founded: 1998 in London
- Headquarters: Berlin

Nuggets tagged by region, checked based on context.

Next Steps

Ready to create nuggets? → Go to Dashboard

Need more details? → Hallucination Detection → API Reference

Want examples? → Knowledge Base Examples