FITC Retrieval-Augmented Generation (RAG) Policy

Version: 1.0 Last updated: April 2026 Effective date: April 23, 2026

1. Purpose

This policy governs the use of Retrieval-Augmented Generation (RAG) within the Friend in the Cloud ("FITC") platform.

Its purpose is to:

• Protect FITC from legal, regulatory, and intellectual property risks
• Define how knowledge-based AI avatars ("Avatars") operate
• Clarify responsibilities between FITC and its users
• Ensure lawful, ethical, and transparent use of indexed content

2. What RAG Means on FITC

Certain Avatars on FITC use Retrieval-Augmented Generation (RAG).

This means:

• Content is stored in indexed form (vector database)
• Relevant information is retrieved at query-time
• AI generates responses using retrieved context

Important distinction: RAG is not purely generative — it involves storage and retrieval of content, which carries specific legal responsibilities.

3. Content Sources & Provenance

FITC commits to indexing only content from the following sources:

Permitted Sources (in order of priority)

• Content owned by FITC
• Content licensed to FITC (explicitly allowing AI/RAG usage)
• User-uploaded content (with legal attestation)
• Public domain materials
• Content with clear open licenses (e.g. Creative Commons, subject to terms)

Strictly Prohibited

• Pirated or unauthorized copyrighted content
• Scraped paywalled or proprietary materials without permission
• Content with unclear or unverifiable licensing
• Confidential or trade-secret materials

FITC maintains the right to remove any content that violates these standards.

4. Nature of Outputs (Critical Protection Layer)

All RAG-generated outputs on FITC are:

• Transformative in nature (summarized, synthesized, interpreted)
• Not intended as verbatim reproduction of source material
• Generated dynamically at query-time

FITC implements safeguards to:

• Reduce verbatim reproduction risk
• Encourage summarization and synthesis
• Detect and limit high similarity outputs

Where appropriate, outputs may include source attribution to:

• Provide transparency
• Support fair-use positioning
• Avoid market substitution concerns

5. User-Uploaded Content (Risk Transfer Layer)

When users upload content to train or customize Avatars:

User Representations

Users confirm that:

• They own the content OR
• They hold valid rights/licenses to use it for AI indexing and retrieval

User Responsibility

Users are solely responsible for:

• Legality of uploaded content
• Any claims arising from that content
• Outputs generated from their training data

Indemnification

Users agree to indemnify and hold FITC harmless against any claims, damages, or legal actions arising from uploaded content.

6. Avatar Responsibility Model

• FITC-managed Avatars: FITC is responsible for content provenance and compliance
• User-created Avatars: The user is responsible for:
  • Training data
  • Output implications
  • Legal compliance

7. Copyright Protection & Takedown Process

FITC operates a copyright complaint and takedown system.

Rights holders may submit requests including:

• Identification of copyrighted work
• Location or description of infringing content
• Proof of ownership
• Contact details and legal declaration

FITC will:

• Review claims in good faith
• Respond within a reasonable timeframe (target: 48 hours to 10 days)
• Remove or restrict access to verified infringing content

FITC reserves the right to:

• Reject incomplete or bad-faith claims
• Request additional verification

8. Personality, Likeness & Avatar Design

To reduce legal exposure:

• FITC does not intentionally replicate real individuals without permission
• Avatars are presented as AI constructs, not real persons
• Named or expert-based Avatars:
  • Require licensing OR
  • Must be sufficiently transformative

Users may not create Avatars that impersonate real individuals without consent.

9. Defamation & Harmful Content

RAG outputs may:

• Combine sources inaccurately
• Present incorrect or misleading information

Therefore:

• Outputs are not guaranteed factual
• Users must independently verify information

FITC reserves the right to:

• Moderate harmful outputs
• Remove content upon valid complaint

10. Data Protection & Privacy (GDPR-Aware)

FITC does not knowingly index:

• Personal data without legal basis
• Sensitive personal information

For user-uploaded data, users must ensure lawful basis (e.g., consent).

Data Deletion

Upon valid request:

• FITC will remove source content
• Remove associated embeddings where reasonably possible

Important limitation: Complete deletion from vector systems may require full system rebuild and may not always be immediately feasible.

11. Prohibited Training Content

Users may not upload:

• Copyrighted material without rights
• Personal data of third parties (without consent)
• Confidential or proprietary business information
• Illegal or harmful content
• Content intended to manipulate or exploit the system

12. No Reliance Clause

All outputs generated via RAG:

• Are probabilistic and may contain errors
• Should not be relied upon for:
  • Financial decisions
  • Legal advice
  • Medical guidance

Use is at the user's own risk.

13. Operational Safeguards (Internal Commitment)

FITC commits to implementing:

• Content provenance tracking
• Output similarity detection mechanisms
• Prompt engineering for transformation (not reproduction)
• Regular audits of indexed content
• Logging of takedown and compliance actions

14. Limitation of Liability

To the maximum extent permitted by law, FITC shall not be liable for:

• User-uploaded content
• Outputs generated from user data
• Indirect, incidental, or consequential damages

15. Policy Evolution

This policy reflects a rapidly evolving legal landscape.

FITC reserves the right to:

• Update this policy at any time
• Adapt to new regulations, case law, and industry standards