Skip to content

Best Practices

A great Knowledge Base is the foundation of effective hallucination detection. These practices help you build one that scales.

Size & Scope

Target: 20-50 Core Facts

Start with essential facts and grow over time:

Phase 1: Foundation (Week 1)

  • 5-10 core facts
  • Company basics: name, founding date, HQ, mission
  • Leadership: CEO name, key roles
  • Product: main product name, tagline
  • Website source for all

Phase 2: Product Launch (Week 2-3)

  • 15-25 facts
  • All product names and pricing
  • Key features and capabilities
  • Customer count or market position
  • Financial basics (funding, ARR)

Phase 3: Maintenance (Week 4+)

  • Add 5-10 facts monthly
  • Update quarterly financials
  • Document new products or partnerships
  • Refine existing facts based on Shield detections

Quality Over Quantity

50 precise, well-sourced facts beat 500 vague claims.

Compare:

Bad set (500 facts, low value):

  • “We’re good at AI”
  • “Our team is smart”
  • “Customers love us”
  • “We’re different from competitors”

Good set (50 facts, high value):

  • “Founded 2024, $15M Series B”
  • “CEO Sarah Chen (Google, Stanford PhD)”
  • “Shield: 94% accuracy, 9+ engines”
  • “500+ enterprise customers”
  • “Only platform with SHA-256 keyed history”

Quality nuggets:

  • Are specific and measurable
  • Have source URLs
  • Reflect current reality
  • Are defensible in audits
  • Produce fewer false Shield alerts

Organization Strategy

Assign Ownership

For teams >5 people, assign nugget categories to owners:

CategoryOwnerReview Frequency
CompanyCEO / Head of OpsQuarterly
ProductsVP ProductMonthly
LeadershipCEO / HRQuarterly
FinancialCFOMonthly (after close)
TechnicalCTOQuarterly
LegalGeneral CounselQuarterly
CompetitiveVP Sales/MarketingMonthly

When facts need updates, owners get reminders. This prevents stale facts from dragging down Shield accuracy.

Document the Why

For each fact, maintain a one-line note on why it’s in the Knowledge Base:

  • “Used in investor pitches” → High importance
  • “Marketing claim, update quarterly” → Needs review before each campaign
  • “Internal benchmark, confidential” → Tag as private
  • “Customer-facing SLA” → Legal holds you to it

Use tags to track this automatically.

Confidence Scoring

Calibration

Confidence scores should reflect reality. Miscalibration creates false Shield alerts:

Over-confident facts (marked 1.0 when actually uncertain)

  • Shield flags minor discrepancies
  • Leads to alert fatigue
  • Teams lose trust in Shield

Under-confident facts (marked 0.6 when well-established)

  • Shield ignores clear hallucinations
  • Misses brand damage
  • Waste of the tool

Guidelines by Category

CategoryTypical ConfidenceExamples
Company0.95-1.0Founded date, HQ, public employee count
Products0.90-0.95Pricing (if public), features (from your docs)
Leadership0.95-1.0CEO/founder names (public)
Financial0.95-1.0 (historic) / 0.80-0.90 (projections)Q4 revenue, Series B amount; Q2 projections
Technical0.85-0.95SLA claims (published), benchmark results (yours)
Legal0.95-1.0Certifications (current), compliance status
Competitive0.70-0.85Market position, relative claims

Red Flags (Mark <0.8)

These belong in your Knowledge Base as lower-confidence claims:

  • “We’re the fastest” (subjective)
  • “Industry-leading features” (marketing)
  • “First to market” (needs verification)
  • Revenue growth rates (estimates)
  • Competitor comparisons (usually disputed)

Shield will still check them, but won’t over-alert.

Maintenance Schedule

Weekly

  • Review Shield alerts pointing to nugget mismatches
  • If nugget needs updating: edit immediately
  • Update prices if you changed them

Monthly

  • Check for new product launches needing nuggets
  • Review team-owned categories for drift
  • Add customer wins or partnerships as relevant

Quarterly

  • Deep review of all facts
  • Update financials (especially quarterly metrics)
  • Archive irrelevant facts
  • Check expiry dates coming up

Annually

  • Full Knowledge Base audit
  • Recalibrate confidence scores based on Shield accuracy
  • Trim outdated facts
  • Plan next 12 months of fact updates

Source Management

Always Include URLs

Every fact should have a source:

Good sources:

  • Company website (/about, /press, /products)
  • Published press releases
  • SEC filings (if public)
  • Published benchmarks or reports
  • Signed contracts (private, internal link)

Avoid:

  • Vague references (“our records show”)
  • Personal claims without backup (“I remember…”)
  • Rumors or industry gossip

Every 6 months, verify source URLs still work:

  1. Go to Knowledge Base → Tools → Check Links
  2. System crawls all source URLs
  3. Reports dead links
  4. Click Fix to update URL or remove source

Time-Sensitive Facts

Expiry Management

Set expiry dates strategically:

Quarterly Financials

  • Fact: “Q4 2025 ARR: $8.2M”
  • Expiry: 2026-03-31 (when Q1 closes)
  • Reminder: 2026-03-24 (1 week before)

Funding Announcements

  • Fact: “Series B: $15M raised March 2024”
  • Expiry: Never (evergreen)

Temporary Claims

  • Fact: “Early-bird pricing: 40% off, limited time”
  • Expiry: 2026-04-30 (when offer ends)

Annual Certifications

  • Fact: “SOC 2 Type II certified through Dec 2025”
  • Expiry: 2025-12-31 (renewal deadline)

Expiry Reminders

You get reminders 1-2 weeks before expiry:

  • Email notification
  • Dashboard badge on Knowledge Base
  • Option to extend, update, or remove

When a fact expires, Shield automatically stops using it. This prevents stale facts from dominating detection.

Handling Shield Detections

When Shield flags an inaccuracy, it might surface a nugget issue:

Scenario 1: Nugget is Wrong

  • AI said: “TruthVouch was founded in 2023”
  • Your nugget: “Founded in 2024” (1.0 confidence)
  • Reality: Actually founded in 2023
  • Fix: Update nugget to “Founded in 2023”, edit reason “Shield detection corrected stale fact”

Scenario 2: AI is Wrong

  • AI said: “TruthVouch costs $10,000/month”
  • Your nugget: “Starting at $349/month” (0.95 confidence)
  • Reality: Your pricing is correct
  • Fix: This is a true positive detection. Deploy correction via Shield.

Scenario 3: Nugget Too Narrow

  • AI said: “TruthVouch offers Compliance AI”
  • Your nugget: “Offers Shield, Brand Intelligence, Governance” (incomplete product list)
  • Reality: You offer 7 products
  • Fix: Update nugget to comprehensive product list

Track these corrections. If Shield frequently detects against a nugget, it’s often a signal the nugget needs updating.

Team Scaling

1-5 People

  • One person (usually CEO) owns all nuggets
  • Update weekly, especially before launches
  • Use Auto-Discovery to reduce manual work

5-20 People

  • Assign category owners (Product, Finance, etc.)
  • Create review schedule
  • Use tags to track status (needs-review, legal-approved)
  • Set edit permissions by category

20+ People

  • Implement workflow: Create → Team Review → Approval → Published
  • Use version history and audit trail
  • Integrate with compliance processes
  • Automate routine updates (daily API calls for uptime SLA)

Integrations

Connect to CMS

Sync facts with your website:

  1. Add facts to Knowledge Base
  2. Enable “Website Sync” in Settings
  3. Facts auto-populate Schema.org JSON-LD
  4. Publish JSON-LD to your domain
  5. AI engines crawl and learn your truth

Connect to Product Docs

Link product facts to your public docs:

Connect to Compliance

Use facts in your Trust Center and public compliance docs:

  • Compliance reports pull from nuggets
  • Certification list auto-syncs
  • Financial summaries pull from Financial nuggets

Avoiding Common Pitfalls

Pitfall 1: Overstuffing

Create one fact per nugget. Don’t combine:

  • “Founded 2024, raised $15M, 150 employees” → 3 nuggets

Pitfall 2: Marketing Inflation

Don’t use marketing language without adjusting confidence:

  • “Industry-leading AI detection” (0.95) → Too confident; should be 0.7-0.8
  • “Enterprise-grade security” (0.90) → Vague; should be 0.6 or define specifically

Pitfall 3: Setting & Forgetting

Knowledge Base needs maintenance:

  • Quarterly financials go stale
  • Product features change
  • Pricing updates

Set reminders and owner assignments.

Pitfall 4: Mixing Public & Private

Don’t mix confidential facts with public ones:

  • Private: “Burn rate $500k/month”
  • Public: “Profitable since Q3 2025”

Use visibility tags (public/private/restricted) to prevent accidental exposure.

Next Steps