Core Concepts
KamerRaad is an AI tool that leverages large language models and hierarchical summarization to enhance citizens' ability to interactively engage with and understand Belgian parliamentary proceedings.
Abstract
KamerRaad is an AI-powered tool developed to improve the accessibility and understanding of Belgian national politics for both expert and non-expert users. The key challenges addressed by KamerRaad include:
The dispersion of political information across various documents and formats, making it difficult to retrieve and contextualize.
The length and density of parliamentary records, which exceeds the context window of large language models (LLMs).
To address these challenges, KamerRaad employs a multi-step approach:
Scraping and chunking of raw parliamentary documents.
Hierarchical summarization, where each document chunk is summarized both comprehensively and concisely. This allows LLMs to process the information within their limited context window.
Metadata tagging of politicians, political parties, and topics to enhance retrieval efficiency.
The KamerRaad user interface, built with Streamlit, facilitates interactive engagement. Users can input queries, which trigger the retrieval of relevant document summaries. They can then request more detailed responses generated by the LLM, which maintain a direct link to the source material through the metadata tags.
By collecting user feedback on the relevance and quality of the responses, KamerRaad aims to continuously improve its source retrieval and summarization, thereby enhancing the overall user experience and promoting deeper understanding of Belgian political discourse.
Stats
Fundamental for democratic participation is that politicians' views, policies, and actions are accessible to citizens.
Parliamentary proceedings are publicly available but vary widely in format and are dense with specialized terminology.
The dispersion of this information across documents from plenary and committee meetings on different days and various types (discussions between politicians vs. proposals of resolutions and laws) further complicates their retrieval and contextualization.
Quotes
"Passing the sources to Large Language Models (LLMs) is a potential solution, but the source documents are too long for even one to fit into the context window of these models."
"KamerRaad manages the challenge posed by extensive parliamentary records, ensuring that users receive precise and contextually relevant information without overwhelming the model's processing capabilities."