Embedding-based retrieval of historical responses for a new Office Action rejection
Reference / educational use to understand domain-tuned embedding models for legal text

Not intended for: direct legal advice, automated filing, or any production attorney-client work without independent practitioner review.

License

CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0).

Research, evaluation, and personal use: permitted with attribution.
Commercial use: requires separate written license. Contact kwisschen@gmail.com for commercial licensing inquiries.

The base model (Qwen/Qwen3-Embedding-0.6B) is Apache-2.0. This adapter and the trained weights are CC-BY-NC-4.0.

Method

Architecture: LoRA on Qwen3-Embedding-0.6B
Trained on US Office Action rejection ↔ applicant response pairs
Contrastive loss with in-batch negatives
Training infrastructure: AWS SageMaker

Limitations

US-prosecution only; not trained on other jurisdictions
Retrieval quality depends on corpus coverage of similar past cases
The model retrieves semantically similar text; it does not assess legal validity or predict outcomes

Disclaimer

This is a research artifact. The Patent-Analyst project is a research demo. Outputs are not legal advice. Any use in actual patent prosecution requires independent review by a licensed practitioner.

Citation

bibtex

@software{chen2026_patent_analyst_embedding,
  author = {Chen, Christopher},
  title  = {Patent-Analyst Embedding LoRA},
  year   = {2026},
}

Commercial inquiries

kwisschen@gmail.com

kwisschen/qwen3-embedding-0.6b-patent-oa-rejections

Patent-Analyst Embedding LoRA

Intended use