Avatar Demo

This folder contains a self-contained copy of the Avatar HotpotQA demo with a FastAPI backend, a lightweight frontend, and the optimized model bundle.

Layout

• server.py – FastAPI app exposing /api/questions and /api/run.
• frontend/ – Static site (index.html, style.css, main.js).
• model/HotpotOptimized – Serialized Avatar agent (copied from the training run).

Prerequisites

Install the Python dependencies (Python 3.10+ recommended):

pip install \
	fastapi "uvicorn[standard]" \
	datasets dspy dill \
	langchain langchain-community \
	arxiv

Notes:

• The backend tries to instantiate LangChain's GoogleSerperAPIWrapper (web search) and ArxivAPIWrapper (paper metadata). If they fail to start—because the packages or credentials are missing—it falls back to an offline cache of the taesiri/arxiv_qa dataset. Installing the packages above enables the real tools.
• To let GoogleSerperAPIWrapper run, export SERPER_API_KEY with a valid Serper.dev token before starting uvicorn.
• The hard-coded DashScope-compatible credentials at the top of server.py mirror the training setup; adjust API_KEY, BASE, and MODEL there (or via environment variables) to match your deployment.

Running the backend

From the Avatar_Demo folder:

uvicorn server:app --host 127.0.0.1 --port 8000 --reload

This serves the API at http://127.0.0.1:8000.

Running the frontend

In a new terminal:

cd frontend
python -m http.server 5500

Then open http://127.0.0.1:5500 in your browser.

Use the dropdown to pick one of the curated questions or type your own, and compare the baseline vs Avatar optimized agent outputs.