1. AI Systems We Use
Retio uses a multi-stage AI pipeline to analyze patent claim structure:
| Stage | Model | Purpose |
|---|---|---|
| Text Encoding | XLM-R Large (550M params, MIT license) | Convert patent claim text into numerical representations |
| Dimensionality Reduction | PCA (1024D → 64D) | Compress representations while preserving structure |
| Structural Embedding | Trained MLP (~7K params) | Map claims into 16-dimensional structural space |
| Similarity Search | pgvector HNSW index | Find structurally nearest patents efficiently |
2. What AI Outputs Mean
Retio outputs are mathematical measurements, not legal conclusions:
- Structural similarity scores measure geometric proximity in embedding space. A high score means two patents have similar claim structures, not that one infringes the other.
- Neighbor lists show patents that are geometrically close. Proximity does not imply any legal relationship.
- Scope metrics (depth, breadth, rank) describe a patent's position in structural space relative to others in its technology domain.
3. What AI Does NOT Do
- Does NOT determine patent infringement
- Does NOT assess patent validity or invalidity
- Does NOT provide freedom-to-operate opinions
- Does NOT generate legal advice of any kind
- Does NOT predict litigation outcomes or damages
- Does NOT replace the judgment of qualified patent professionals
4. Training Data
Our models are trained exclusively on publicly available patent data from government patent offices:
- USPTO (United States Patent and Trademark Office)
- KIPRIS / KIPO (Korean Intellectual Property Office)
- JPO (Japan Patent Office)
- EPO (European Patent Office)
We do not train models on user content. User search queries, generated reports, account data, and monitoring configurations are never used for model training or improvement.
5. Limitations & Known Risks
- False positives: The system may identify structural similarity where no meaningful legal relationship exists
- False negatives: The system may miss structurally relevant patents, particularly in domains with sparse training data
- Language bias: The model was primarily trained on Korean patent data with cross-lingual transfer to English. Coverage of other languages may be limited.
- Temporal gap: Newly published patents require embedding computation before they appear in search results (typically within 1 week)
- Dimensional compression: Mapping complex claim structures into 16 dimensions necessarily loses some information
6. Human Oversight
Retio is designed as a decision-support tool, not an autonomous decision-maker:
- All outputs require human interpretation by qualified professionals
- No automated actions are taken based on similarity scores (e.g., no automated demand letters)
- Monitoring alerts are informational — they notify users of potential structural overlap for manual review
- Users must independently verify any findings before taking legal or business action
7. Regulatory Compliance
Korean AI Basic Act (AI 기본법, 제31조)
본 서비스의 모든 분석 결과는 인공지능 알고리즘에 의해 생성됩니다. 모든 AI 생성 출력물에는 AI 생성 표시가 부착됩니다.
EU AI Act (Regulation 2024/1689, Art. 50)
In compliance with EU AI Act transparency obligations, we disclose that all analytical outputs are generated by AI systems. All AI-generated content is labeled with an AI disclosure badge.
8. AI Disclosure on Outputs
Every Retio output — reports, search results, monitoring alerts, and exported documents — includes a visible AI disclosure badge:
AI-generated structural analysis. Not legal advice.
This badge may not be removed, obscured, or modified when sharing or exporting Retio outputs.
9. Contact
For questions about our AI systems or this policy, contact legal@retio.ai.