# Import `spaces` first — must happen before any torch/CUDA initialization # (torch is pulled in transitively via rag_chain → langchain_huggingface). # The try/except keeps local dev working (spaces is HF-Spaces-only). try: import spaces # noqa: F401 — imported for GPU-context registration, not called directly except ImportError: pass import gradio as gr from rag_chain import load_chain_components, answer_question, init_gpu # Register GPU entry point with ZeroGPU runtime (satisfies the # "No @spaces.GPU function detected" check). Must run before # load_chain_components. Returns a plain bool — picklable across the # ZeroGPU worker-process boundary. init_gpu() # Load index + LLM once at startup (not per-request) vectorstore, llm = load_chain_components() SAMPLE_QUESTIONS = [ "What do Kaplan et al. and the Chinchilla paper disagree about?", "How does the RAG paper combine retrieval with generation?", "Why do language models hallucinate, according to the sources?", "How was the disagreement between Kaplan and Chinchilla resolved?", ] def respond(question, history): if not question or not question.strip(): return history, "" answer_text, sources = answer_question(question, vectorstore, llm) # Format the citation key as a collapsible-looking block under the answer citation_block = "\n\n---\n**Sources**\n" for s in sources: citation_block += f"\n**[{s['n']}]** {s['citation']} — p.{s['page']}\n> {s['excerpt']}\n" full_response = answer_text + citation_block # Gradio 6.x Chatbot expects messages as dicts with 'role' and 'content' # keys (not (user, bot) tuples as in Gradio 5.x). Initial history may be None. if history is None: history = [] history = history + [ {"role": "user", "content": question}, {"role": "assistant", "content": full_response}, ] return history, "" with gr.Blocks(title="Groundwork") as demo: gr.Markdown( """ # Groundwork Ask a question spanning multiple GenAI research papers. Every claim in the answer is cited back to the exact source excerpt it came from. **Corpus:** Kaplan et al. 2020 (Scaling Laws) · Hoffmann et al. 2022 (Chinchilla) · Pearce & Song 2024 (Reconciling) · Lewis et al. 2020 (RAG) · Kalai et al. 2025 (Hallucination) """ ) chatbot = gr.Chatbot(label="Groundwork", height=450) msg = gr.Textbox( label="Your question", placeholder="e.g. What do Kaplan et al. and Chinchilla disagree about?", ) with gr.Row(): submit_btn = gr.Button("Ask", variant="primary") clear_btn = gr.ClearButton([msg, chatbot], value="Clear") gr.Examples(examples=SAMPLE_QUESTIONS, inputs=msg, label="Try one of these") msg.submit(respond, [msg, chatbot], [chatbot, msg]) submit_btn.click(respond, [msg, chatbot], [chatbot, msg]) if __name__ == "__main__": demo.launch()