Chicago Budget RAG

This project is a public-facing RAG application for the Chicago FY2026 budget documents. It lets users ask plain-English questions about the Annual Appropriation Ordinance and the Grant Details Ordinance, then returns answers with page-level citations and direct links back to the source PDFs.

The system is designed around three goals:

• retrieval quality for long civic PDFs
• transparent source grounding with page citations
• practical deployment for public use

What It Does

• extracts text from the two source PDFs
• chunks and indexes the content with page metadata
• blends BM25 and optional vector retrieval
• optionally reranks results with a cross-encoder
• generates cited answers using OpenAI, Bedrock, or Ollama
• lets users open exact source pages in a built-in viewer
• supports exporting answers to Markdown, JSON, and CSV
• includes evaluation tooling and tuning scripts for retrieval quality

Source Documents

• chicago_Annual_Appropriation_Ordinance_2026.pdf
• chicago_Grant_Details_Ordinance_2026.pdf

Project Structure

• app.py: FastAPI app, routes, SEO pages, export endpoints, analytics hooks, runtime controls
• src/chicago_budget_rag/engine.py: indexing, retrieval, reranking, answer generation, provider abstraction
• templates/: HTML templates for search, guides, disabled page, and rate-limit page
• static/styles.css: shared styling for the web UI
• build_index.py: offline index build entrypoint
• query_rag.py: CLI query tool
• eval_rag.py: retrieval evaluation and tuning harness
• eval/questions.sample.json: starter benchmark set
• docker-compose.yml: containerized runtime configuration
• Dockerfile: image build definition
• .env.openai.example: OpenAI-based runtime template
• .env.bedrock.example: Bedrock-based runtime template for Lightsail or other AWS hosts

Documentation

• Local launch: docs/local/README.md
• AWS launch: docs/aws/README.md

License

This project is released under the MIT License. See LICENSE.

Runtime Notes

• POST / redirects to canonical GET /search?q=... URLs for crawlable search pages.
• Curated search pages and guide pages are indexable; arbitrary search pages are noindex,follow.
• robots.txt and sitemap.xml are served by the app.
• The public site can be disabled with SITE_ENABLED=false while keeping health checks alive.
• Query export is available through the UI and GET /export.

Model Providers

The app supports:

• OpenAI
• AWS Bedrock
• Ollama

Provider selection and model configuration are environment-driven.

SEO Surface

The app includes:

• canonical search result pages
• guide landing pages
• robots.txt
• sitemap.xml
• Open Graph and Twitter metadata
• JSON-LD for home, search, and guide pages
• internal linking through guides and curated searches

Evaluation

Use eval_rag.py to measure hit rate and MRR across a benchmark set and tune BM25/vector blend settings before shipping retrieval changes.