Bloomington Tourist Assistant

A deployed RAG service for Bloomington, Indiana tourism questions. The system combines vector search, BM25, reciprocal rank fusion, cross-encoder reranking, LangGraph routing, and evaluation checks that run in GitHub Actions.

Live app: https://bloomington-rag-473936713859.us-central1.run.app/

Architecture

flowchart TD
    U["User question"] --> A["FastAPI /ask endpoint"]
    A --> R["LangGraph router"]
    R -->|simple| S["Single retrieval path"]
    R -->|complex| D["Query decomposition"]
    S --> V["Qdrant vector search"]
    D --> V
    S --> B["BM25 keyword search"]
    D --> B
    V --> F["Reciprocal rank fusion"]
    B --> F
    F --> C["SBERT cross-encoder reranker"]
    C --> G["Groq LLM synthesis"]
    G --> UI["Chat UI response"]

Loading

Proof points

Area	Evidence
Retrieval quality	86.67% top-1 retrieval accuracy, 13/15 eval questions
CI gate	GitHub Actions runs retrieval tests on every push and pull request
Faithfulness check	LLM-as-judge scored hotel queries at 0.89
Deployment	GCP Cloud Run with a FastAPI backend and static chat UI
Known tradeoff	In-memory Qdrant rebuilds on cold start

Tech stack

• FastAPI, Uvicorn
• LangChain, LangGraph
• Qdrant in-memory vector store
• BM25, reciprocal rank fusion, SBERT reranking
• Groq llama-3.1-8b-instant
• GitHub Actions
• GCP Cloud Run

Setup

git clone https://github.com/prathima-sola/bloomington-rag
cd bloomington-rag
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
python src/prepare_data.py
uvicorn src.api:app --reload --port 8000

Open http://localhost:8000.

Environment

GROQ_API_KEY=

The retrieval quality test does not need an API key. The live answer generation path needs GROQ_API_KEY.

Test commands

cd tests
python test_retrieval.py
python faithfulness_eval.py

test_retrieval.py runs without LLM API calls. faithfulness_eval.py uses Groq.

Deployment notes

• Procfile starts the FastAPI app for Cloud Run.
• The app loads processed documents, builds the in-memory Qdrant index, and serves the chat UI from src/static.
• Cold starts rebuild the vector index. A managed Qdrant instance would remove that rebuild time.

Project structure

bloomington-rag/
├── src/
│   ├── api.py
│   ├── rag_pipeline.py
│   ├── agentic_rag.py
│   ├── prepare_data.py
│   └── static/index.html
├── data/processed/
├── tests/
│   ├── test_retrieval.py
│   ├── faithfulness_eval.py
│   └── eval_dataset.py
├── .github/workflows/quality_gate.yml
└── Procfile

Known limits

• The dataset is heavier on hotels than restaurants and attractions.
• Qdrant runs in memory for this version.
• The source data is static, so events, hours, and prices can age.