| Introduction

Government organizations often face significant challenges in processing large volumes of legislative bills. Manual tasks such as data collection, verification, and summarization can delay critical insights, slowing down timely decision-making. Metrum AI offers a comprehensive solution to these challenges by leveraging small language models (SLMs) and agentic retrieval-augmented generation (RAG) techniques to streamline legislative bill analysis. This solution harnesses the power of AI agents—intelligent systems capable of operating semi-autonomously to complete tasks like summarization and data validation in offline, batch modes. Deployed on the latest Dell PowerEdge servers featuring 5th Gen AMD EPYC 9755 128-Core processors, this solution ensures both scalability and robust performance.

By running entirely on CPUs, Metrum AI's solution allows organizations to leverage their existing CPU-based infrastructure, eliminating the need for costly specialized hardware.

| Key Highlights

| Solution Overview

Metrum AI's legislative assistant simplifies legislative bill analysis by leveraging specialized AI agents to automate key tasks such as evaluating economic impact, legal compliance, and social or environmental effects. By combining AI-driven efficiency with human oversight, analysts can review findings, make adjustments, and ensure the creation of accurate and comprehensive reports.

Figure 1. User Interface of the Legislative Bill Analysis Solution.

Figure 2. vLLM Model Serving Performance of Llama 3.2 3B with BF16 Precision

Figure 3. Performance results for legislative bill analysis solution.

This illustrates an 11.69x increase in throughput with only a 2.57x increase in processing time when comparing the analysis of 1 document to 32 documents. This highlights the efficiency of AI agents in executing batched tasks such as document summarization.

To support this solution, we chose the Dell PowerEdge R7725 server equipped with 5th Gen AMD EPYC 9755 128-Core processors and high-speed DDR5 6000 memory.

| Solution Details and Workflow

Figure 5: Solution Workflow.

Bill Upload and Preprocessing: Users begin by uploading draft bills or legislative documents into the system.
Contextual Retrieval via Vector Database: The system leverages a vector database (VectorDB) to store and retrieve legislative context.
Agentic RAG Analysis: AI agents are deployed to focus on various dimensions of the bill.
Comprehensive Report Generation: Once all agents complete their tasks, the system compiles findings and generates a detailed report.

| Solution Architecture

Figure 6. Solution Architecture.

The software stack incorporates the following key components:

vLLM (v0.5.3.post1): An industry-standard library for optimized serving of open-source large language models (LLMs).
llama-deploy: An async-first framework designed for building, iterating, and deploying multi-agent systems.
Llama 3.2 3B Model: A leading open-weight small language model with three billion parameters.
LlamaIndex: A widely-used open-source retrieval-augmented generation (RAG) framework.
bge-large-en Embeddings Model: A top-ranked text embeddings model.
MilvusDB: An open-source vector database offering high-performance embedding and similarity search capabilities.

| Summary

This implementation demonstrates how government organizations can overcome the challenges of processing large volumes of legislative bills by utilizing small language models (SLMs) and agentic retrieval-augmented generation (RAG) techniques.

"The advantage of AI agents presents an unparalleled productivity opportunity, enabling AI to operate independently to accomplish tasks with minimal human intervention."

Chetan Gadgil, CTO at Metrum AI

To learn more about this solution, check out our repository: https://github.com/metrum-ai/genai-legislative-insights.

Copyright © 2025 Metrum AI Inc. All Rights Reserved. Dell Technologies, Dell, Dell PowerEdge, Dell logo, and other trademarks are trademarks of Dell Inc. or its subsidiaries. AMD, AMD logo, AMD EPYC, AMD ROCm, and combinations thereof are trademarks of Advanced Micro Devices, Inc.