Neural Fact Sheets

Neural Fact Sheets are AI-friendly summaries of your facts, designed to help LLMs generate correct responses. Learn how to create effective fact sheets that minimize hallucinations.

What Is a Neural Fact Sheet?

A Neural Fact Sheet is a structured document containing a fact, its context, examples, and source. It’s converted to vector embeddings and stored in a database so that when an LLM is queried, relevant facts are automatically injected into its context.

Structure

FACT ID: fact_pricing_standard
FACT TITLE: TruthVouch Standard Plan Pricing
STATEMENT: TruthVouch Standard plan costs $500 per month
CATEGORY: Pricing
CONTEXT:
  Effective from Q1 2024. The Standard plan includes up to 5 million
  cross-checks per month and supports up to 3 Truth Nuggets. Upgrades
  to Professional or Enterprise available for higher volumes.
SOURCE:
  - Primary: pricing.truthvouch.com (updated 2024-01-15)
  - Backup: Internal pricing policy document
CONFIDENCE: High
LAST UPDATED: 2024-01-15
EXAMPLES:
  - "The Standard plan costs $500 per month"
  - "For $500/month, you get the Standard plan"
  - "$500 monthly subscription for Standard features"
  - "Pricing starts at $500/month for the Standard tier"
RELATED FACTS:
  - Professional plan: $1,500/month
  - Enterprise: Custom pricing
  - Volume discounts: Available above 20M checks/month

Creating Effective Fact Sheets

1. Clear, Concise Statement

The STATEMENT is the core fact — make it unambiguous:

Weak:

"TruthVouch pricing is $500"

Better:

"TruthVouch Standard plan costs $500 per month"

Why better?

Specifies plan (Standard, not others)
Specifies period (monthly)
Avoids ambiguity

2. Rich Context

The CONTEXT helps the AI understand nuance and avoid false inferences:

Weak context:

"Effective Q1 2024"

Rich context:

"Effective from Q1 2024 (January 2024). The Standard plan is our mid-tier offering,
designed for teams with up to 50 Truth Nuggets. It includes up to 5 million cross-checks
per month. For higher volumes, see Professional ($1,500/month) or Enterprise (custom pricing).
Volume discounts available for organizations exceeding 20M checks/month."

Why rich context helps?

Clarifies what Standard plan is for
Mentions tiers (prevents hallucination of intermediate prices)
Explains volume discount eligibility
Provides reasoning (helps LLM generalize correctly)

3. Concrete Examples

EXAMPLES show the LLM different ways to phrase the fact:

Weak examples:

- "$500/month"
- "Standard costs $500"

Better examples:

- "The TruthVouch Standard plan costs $500 per month"
- "$500 per month gets you our Standard plan"
- "For $500/month, you'll have access to 5M monthly cross-checks"
- "TruthVouch's Standard tier is $500 monthly"
- "The Standard offering from TruthVouch is $500 a month"

Why multiple examples help?

Shows variation in phrasing (AI learns idioms)
Prevents rigid template-matching
Covers synonyms (“monthly” = “per month”)
Helps with different contexts (question vs. statement vs. comparison)

4. Credible Source

SOURCE tells the AI where this fact comes from (authority):

Weak source:

"We think this is the price"

Strong sources:

- Primary: https://pricing.truthvouch.com (official pricing page, updated 2024-01-15)
- Backup: Internal pricing policy (confidential document, version 2.3)
- Authority: CFO approved 2024-01-10

Why source matters?

High-authority sources → Higher confidence
Official/public sources → LLM more likely to cite correctly
Version dates → Help identify outdated information

5. Confidence Level

CONFIDENCE indicates certainty (affects LLM behavior):

Level	Definition	Use When
High	Definitive, official source	Published pricing, official announcements
Medium	Reliable but not official	Internal docs, team consensus
Low	Approximate or uncertain	Estimates, “likely to be…”, pending confirmation

Bad confidence: “Medium (we’re pretty sure)”

Good confidence: “High (from official pricing page)“

RELATED FACTS prevent isolated hallucinations:

Without related facts:

AI might hallucinate intermediate pricing (“$700/month plan exists”)

With related facts:

RELATED FACTS:
  - Professional plan: $1,500/month (5x Standard)
  - Enterprise plan: Custom pricing (contact sales)
  - Annual discount: 15% off when paying yearly
  - Volume discounts: Available above 20M checks/month

Why helpful?

Gives AI sense of your pricing structure
Prevents invention of non-existent tiers
Helps with comparative questions (“Which plan is best for X?”)

Fact Sheet Best Practices

Do

Be specific: “TruthVouch’s Shield product” (not “our product”)
Use official terminology: Match what you’d say on your website
Provide range context: If price varies, explain what affects variation
Update regularly: Monthly or quarterly, especially for dynamic facts
Link to sources: Make sources verifiable and accessible
Use current information: Outdated facts cause hallucinations

Don’t

Don’t be verbose: Context should be 2-3 sentences, not paragraphs
Don’t include marketing hype: Stick to facts (“best-in-class” is vague)
Don’t mix facts: One fact sheet per fact (not “pricing and features”)
Don’t include outdated versions: Archive old versions separately
Don’t rely on AI to interpret: Be explicit (“$500/month” not “competitive pricing”)

Organizing Fact Sheets

By Category

Organize related facts together:

Financial Facts:

Pricing plans (Standard, Professional, Enterprise)
Discount policies
Billing terms
Revenue metrics

Product Facts:

Feature availability
Integration compatibility
System requirements
Update cadence

Company Facts:

Founded date
Employee count (approximate)
Office locations
Leadership team

Legal Facts:

Certifications (SOC 2, ISO, etc.)
Compliance claims (HIPAA-ready, GDPR-compliant)
Data residency options
Support SLA

Versioning

Keep version history:

Current: v2.1 (2024-03-15)
Previous: v2.0 (2024-01-15)
Previous: v1.5 (2023-11-01)

This helps with:

Auditing changes
Reverting if update was wrong
Understanding evolution of facts
Compliance (document change history)

Testing Fact Sheet Effectiveness

A/B Testing

Test two versions of a fact sheet:

Variant A (Current):

STATEMENT: TruthVouch Standard plan costs $500 per month
CONTEXT: Mid-tier offering with 5M checks/month
EXAMPLES:
  - "$500/month for Standard"
  - "Standard plan is $500/month"

Variant B (New):

STATEMENT: TruthVouch Standard plan costs $500 per month
CONTEXT: Mid-tier offering with 5M checks/month. Designed for teams
  with 1-50 Truth Nuggets. Professional tier starts at $1,500/month.
EXAMPLES:
  - "$500/month for Standard"
  - "Standard plan is $500/month"
  - "For $500, get 5M checks with Standard"
  - "Standard: $500/month, Professional: $1,500/month"

Evaluation:

Deploy Variant B to vector DB
Monitor hallucinations for 1 week
If Variant B performs better → Replace Variant A
Track metric: % of queries answering pricing correctly

Quality Scoring

Rate your fact sheets on quality:

Dimension	Poor	Fair	Good
Clarity	Ambiguous phrasing	Some ambiguity	Crystal clear
Context	Missing context	Minimal context	Rich context
Examples	0-1 examples	2-3 examples	4+ variations
Sourcing	No source	Generic source	Official source with date
Currency	Outdated	Possibly outdated	Recently updated
Completeness	Isolated	Some relations	Full relationship map

Target: Good on all dimensions

Maintenance & Updates

When to Update

Update immediately:

Pricing changes
New product releases
Compliance/legal changes
High hallucination rate (>20% of queries wrong)

Update quarterly:

Employee count
Office locations
Leadership changes
General refresh

Update annually:

Awards/certifications
Revenue metrics
Partnership announcements

Update Process

Identify need: Which fact needs updating?
Draft update: New version of fact sheet
Verify source: Confirm accuracy with authoritative source
Test variant: A/B test new version if possible
Deploy: Replace old version in vector DB
Archive: Keep old version in history
Monitor: Track effectiveness for 1 week

Auto-Correction — How fact sheets are deployed as corrections
Correction History — Track fact sheet updates and effectiveness
Overview — Corrections overview

Next Steps

List critical facts — Which facts must be accurate?
Create fact sheets — Draft for top 10 facts
Test quality — Rate on 6 dimensions above
Deploy — Add to vector DB / knowledge base
Monitor effectiveness — Track hallucination rate
Iterate — Improve fact sheets based on real-world performance