VoteWise — Belgian Political Program Comparator

Overview

VoteWise is a Retrieval-Augmented Generation (RAG) system designed to help users explore, summarize, and compare political party positions in Belgium. It combines Azure AI Search for document indexing and vector search with Azure AI Foundry for embeddings and language model reasoning. Users can ask natural-language questions about party programs, and the system retrieves the most relevant information before generating precise, context-aware answers.

Architecture

Key Components

Document Storage & Indexing

• Documents are preprocessed, split into chunks, and indexed in Azure Cognitive Search with vector embeddings.
• Vector search allows efficient retrieval of the most relevant document chunks for a given query.

Embedding & Language Model

• Azure AI Foundry embeddings are used to represent each document chunk in vector space.
• Azure AI Foundry LLMs generate context-aware summaries and answers.
• Filtering by party ensures results are relevant to the user’s query.

API Layer

• FastAPI serves as the REST API endpoint (/rag), receiving queries as JSON.
• Returns both the model-generated answer and the retrieved context chunks with source metadata.
• Example query for testing in Swagger UI:

{
  "question": "Quels sont les points clés du programme de l'Ecolo pour l'environnement ?",
  "party": "ecolo"
}

Deployment on Azure (in progress...)

• Party programs, news articles, and other political documents will be stored in Azure Blob Storage.
• Web App for Containers will host the FastAPI RAG service and/or a streamlit app.
• Azure Container Registry (ACR) will store the container images.
• Managed Identity with AcrPull role will allow secure container pull.
• CI/CD pipeline will deploy new container versions automatically via GitHub Actions.
• Logs and monitoring through Azure Log Analytics.

Usage

Local Docker Testing

docker build -t votewise-rag:latest .
docker run -p 8000:8000 --env-file .env votewise-rag:latest

• Access http://localhost:8000/docs for Swagger UI.
• Use example queries to test RAG retrieval.

Possible Improvements & Next Steps

• Integrate automated scraping from RTBF news to enrich document corpus with up-to-date political articles.
• Add social media streams (e.g., X/Twitter) for party mentions.
• Implement comparison between party programs and actual government measures.
• Extend language support for Flemish parties (Dutch).
• Test RAG performance.
• Improve document ingestion (PDF loading) and chunking.
• Introduce unit testing, CI/CD pipelines, and scheduled updates for automated indexing.

Notes on Filtering

• Filtering by party ensures that only relevant documents are retrieved.
• Example: querying MR pension proposals will only retrieve MR-related content.

Source chunk used for question on MR pension without filtering activated:

travail des mesures pour les fins de carrière. Le PS propose de : ... (Source: ps-federal-2024.pdf)

Free Tier Limitations

Using Azure free tiers for VoteWise RAG is fine for prototyping, but comes with constraints: limited Cognitive Search index size (50MB) and query throughput, restricted OpenAI/Foundry requests and models, single-instance Web App with no auto-scaling, and small Blob storage capacity. Free tiers are not suitable for large datasets, real-time monitoring, or production workloads.