从混乱到清晰：Mary Technology 如何在 AWS 上构建用于智能体 AI 的法律事实层.pdf

1、S T P 1 0 6-SFrom chaos to clarity:How Mary Technology is building the legal Fact LayerDaniel Lord-DoyleCEO&Co-Founder,Mary Technology 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.LLMs,even with Retrieval-Augmented Generation or agentic frameworks,are not fit-for-purpose for le

2、gal dispute-resolution workloads.The Problem 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.P R O B L E M#1The availability of training dataThere isnt a single“right answer”The Training Problem 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.P R O B L E M#2LLMs

3、 are compression machines 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.OCR text(pixels characters).Text tokens(sequence compression).Tokens embeddings(semantic compression).Attention layers(contextual compression).Long-doc strategies(arbitrary chunking,summarisation,retrieval).

4、Example stages of compression 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.LLMs are built to:Handle a massive range of tasksScale across billions of documents and trillions of tokens efficientlyGenerate fluent,plausible text without deep pre-processingWork without being retrain

5、ed for every new task or domainThey are generalists-and thats their strength.*This slide was written by an LLM.Can you tell?2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.P R O B L E M#3Facts are not in the uploaded data 2025,Amazon Web Services,Inc.or its affiliates.All rights r

6、eserved.Ambiguous:If both Alice Smith and Andrew Smith exist in the documents,which one is it?Date:Is that 5th March or 3rd May?Context:In this matter,is a reported error on this date potentially important?What are the error details?Fragmented:Do other documents mention this fact?Do they support or