By utilizing a RAG framework, the system bypassed the need for fine-tuning the LLM, ensuring scalability and adaptability to complex legal requirements without additional model training.
"Automated Legal Analysis of Financial Transaction Agreements using GPT, Vector RAG, Parameterized Meta Prompting"
The Need
A global investment bank required a solution to migrate 1,800 derivative trades into a new system. Each trade involved a 20-page legal agreement, spanning multiple jurisdictions and subject to complex legal statutes and precedents. With fewer than 600 classified documents available for training, manual processing was infeasible, presenting significant risks in scalability, compliance, and efficiency.
The Approach
Atyeti leveraged a pre-trained large language model (LLM) in conjunction with a vector retrieval-augmented generation (RAG) framework. By integrating legal guidelines, laws, and precedents into the RAG system, the solution ensured accurate extraction and categorization of critical legal details. Parameterized meta prompt engineering enabled precise alignment with jurisdictional requirements, facilitating the automated classification of each agreement.
The Process:
- Identifying and organizing relevant legal statutes and guidelines for incorporation into the RAG system.
- Building a robust reference database to ensure comprehensive and jurisdiction-specific analysis.
- Collaborating with legal SMEs to validate the outputs for accuracy and relevance.
Result
The automated system successfully classified and processed all 1,800 agreements without human intervention. By overcoming the limitations of minimal training data, the solution provided an efficient and compliant method to handling large unstructured data and extracting critical information for onboarding into the new system.