This Space is running in a free, CPU-only environment. That means:

• Higher latency is expected for model-heavy steps (CL encoding, property prediction, and inverse design).
• Cold starts can occur after inactivity (the container spins down), so the first request may take longer.
• Some operations are compute-bound on CPU (DeBERTav2,Transformer/GINE/SchNet encoders), so throughput is limited compared to GPU.

Scaling note: If usage grows, this deployment can be migrated to a GPU-backed runtime (and/or a queued worker setup) to reduce per-request latency and improve concurrency. The app is designed to be hardware-agnostic—the same workflow runs on CPU today and can be accelerated on GPU later with minimal code changes.

How to use PolyAgent

PolyAgent is a web app with three Tabs at the top:

• PolyAgent Console (main workflow)
• Tools (run individual tools)
• Other LLMs (baseline LLM-only answers)

PolyAgent Console

Use this Tab for the full, end-to-end run.

1. In Questions, paste your request (one question or multiple).
2. Click Run PolyAgent.
3. Read the results in:
   • PolyAgent Answer: the final structured response.
   • PNG Artifacts: any available visuals (molecule render, generation grid, explainability heatmap).

Prompt tips (what PolyAgent detects automatically):

• Inverse design / generation: include words like generate or inverse design and include a numeric target (examples: target_value=60, target: 60, Tg 60).
• Seed polymer: provide a pSMILES either:
  • inside a fenced code block, or
  • with a keyed prefix like seed_psmiles:.
• Citations: if you want a specific count, say it explicitly (example: cite 10 papers).

Tools (debugging / run one step at a time)

Use this Tab when you want to run a single tool and inspect its raw output. Each section is a collapsible Accordion with its own inputs and a run button:

• Data Extraction (parse/canonicalize pSMILES; may also produce PNGs)
• Property Prediction
• Polymer Generation (inverse design)
• Web / RAG (search + retrieval)
• Explainability
• Diagnostics (health checks, e.g., OpenAI probe)

Outputs appear as JSON (for tool results) and/or PNGs (for visuals), depending on the tool.

Other LLMs (no tools)

Use this Tab to get a direct answer from a selected non-GPT model. It does not run PolyAgent tools (no property prediction, no generation tools, no retrieval). Pick a model, paste your prompt, and run it.