product updateAmazon Web Services

AWS Publishes Reference Architecture for Contract Intelligence Using Bedrock AgentCore and Dual Claude Models

TL;DR

AWS published a reference architecture showing how to combine structured data extraction with Bedrock AgentCore, dual Claude models, and Amazon Quick to answer portfolio-wide questions that standard RAG systems get wrong. The design uses Claude Sonnet 4.6 for extraction and Claude Haiku 4.5 for independent verification, with Amazon Textract as a deterministic tiebreaker.

3 min read
0

AWS has published a detailed reference architecture for a contract intelligence platform that addresses a specific failure mode in Retrieval Augmented Generation (RAG) systems: aggregation queries across large document sets. The post, published on the AWS Machine Learning blog, walks through a working system that combines Amazon Bedrock AgentCore, the open-source Strands Agents SDK, Amazon Quick embedded dashboards, and two different Claude models running in a verification loop.

The core problem: RAG-based chat tools chunk documents into vector embeddings and retrieve only the top-k most relevant chunks to answer a question. This works well for single-document lookups ("What are the payment terms in this contract?") but fails on portfolio-wide aggregation ("What's our total contract value across 250 vendor contracts?"). Because the full dataset never reaches the model's context, RAG systems return confident but incorrect totals when asked to sum, count, or compare across many documents.

AWS's proposed fix is architectural, not prompt-based: extract structured fields from unstructured PDFs into a relational database, then use analytics tools built for aggregation, while keeping a separate knowledge base for single-document semantic search.

Pipeline design

According to AWS, contracts uploaded to an Amazon S3 bucket trigger an automated pipeline with these stages:

  • An extraction agent, powered by what AWS describes as Claude Sonnet 4.6, reads each PDF natively (no OCR preprocessing) and extracts eight key fields with confidence scores, returned as structured JSON.
  • A separate verification agent, powered by Claude Haiku 4.5, independently re-reads the same document and checks the extraction.
  • If the two agents disagree on signature detection specifically, Amazon Textract's computer vision provides a deterministic tiebreaker.
  • Verified results are stored in Amazon Aurora PostgreSQL.
  • Users query the data through Amazon Quick embedded dashboards and a natural language chat agent, with pipeline status streamed over WebSocket in real time.

Both agents are built on the Strands Agents SDK and deployed on the AgentCore runtime within Amazon Bedrock AgentCore, which AWS says handles serverless hosting, automatic scaling, and session isolation without requiring infrastructure management. Access control is enforced through Cedar-based policies in AgentCore's Policy feature, which AWS states can be validated with automated reasoning before enforcement — intended to restrict which contracts agents and users can access given the sensitivity of pricing and vendor data.

AWS says the dual-model design is deliberate: using two different foundation models with different training data for extraction and verification is meant to catch errors that re-running the same model would miss, since a single model can hallucinate a value with high confidence without any independent check. AWS states the pipeline can process a contract "in seconds under typical conditions" and is designed to scale to handle many contracts in parallel, though no specific throughput or accuracy benchmarks were disclosed.

What this means

This is not a new model launch — it's an architecture pattern for enterprises hitting RAG's known ceiling on aggregation queries over large document corpora. The notable technical detail is the use of two distinct Claude model tiers (a larger extraction model and a smaller, faster verification model) as a cross-check mechanism rather than relying on self-consistency from a single model. For teams building document-processing pipelines on Bedrock, the pattern — extract to structured storage for math, retrieve from a knowledge base for lookups — is a reusable blueprint beyond contracts, applicable to invoices, compliance filings, or any corpus where leadership asks portfolio-level questions that chunk-based retrieval cannot answer.

Related Articles

product update

Replit brings Microsoft Fabric, Teams and local work to general availability; Azure publishing in early access

At Microsoft Ignite, Replit announced general availability for its Microsoft Fabric integration, local access in Replit Desktop, and a Microsoft Teams integration. Publishing apps to a company's Azure tenant is in early access, with Full BYOC planned next. Replit was also named 2026 Microsoft Partner of the Year for Startups.

product update

Perplexity Brain Can Now Create and Update Agent Skills; 80,000+ Personal Skills Written for 20,000+ Users

Perplexity said on November 18, 2026, that Brain, the personalization system for its Computer agent, can now create and update reusable agent skills. According to the company, Brain wrote more than 80,000 personal skills for over 20,000 users between August 1, 2026, and November 14, 2026.

product update

Mistral launches Vibe Desktop for macOS and Windows, running parallel agent tasks locally or in the cloud

Mistral has released Vibe Desktop, a desktop app for its Vibe agent (formerly Le Chat), for macOS and Windows. It runs multiple tasks in parallel in a local repository or a cloud sandbox, and a Linux version is planned.

product update

Microsoft launches Autopilot agent, Home interface and Copilotforce with Salesforce at Ignite 2026

At Ignite 2026, Microsoft introduced Autopilot, an always-on Copilot agent with its own Entra identity, available now to Frontier program subscribers. It also unveiled a new Home interface, Copilot Workspaces, Copilot Code and Copilotforce, a Salesforce integration with 38 prebuilt skills.

Comments

Loading...