ai-scientist-v2

name

ai-scientist-v2

description

Operate and troubleshoot the local AI-Scientist-v2 / AI Scientist Cloud project at /home/qiao/dockerai/AI-Scientist-v2. Use when the user asks to run AI-Scientist-v2, generate or manage research ideas, launch BFTS paper-generation experiments, configure OpenAI/Semantic Scholar keys, use the Streamlit AI Scientist Cloud console, inspect experiment outputs/logs/PDFs/reviews, debug this project’s launcher/web app/configuration, or run paper queries/citation searches that should use the s2-api skill with S2_API_KEY from /home/qiao/dockerai/AI-Scientist-v2/.env.

AI-Scientist-v2

Project root

Default local project: /home/qiao/dockerai/AI-Scientist-v2.

Always run commands from that directory unless the user gives a different checkout. Treat this project as an autonomous code-execution system: prefer sandboxed/container execution, avoid exposing secrets, and do not run expensive experiments without confirming model, idea file, cost/timeout, and whether writeup/review should run.

For detailed command recipes and file map, read references/project-guide.md.

Fast workflow

1. Check readiness

   • Verify .env or environment variables for OPENAI_API_KEY; use S2_API_KEY for higher-throughput Semantic Scholar literature search.
   • Use AI_SCIENTIST_MODEL as the default model for ideation when --model is omitted.
   • Inspect bfts_config.yaml before long runs; the effective BFTS knobs live under agent, agent.stages, agent.multi_seed_eval, and agent.search.

2. Generate ideas

   • Input is a topic/workshop Markdown file with title/keywords/TL;DR/abstract-style scope.
   • Run ai_scientist/perform_ideation_temp_free.py; it writes a same-stem .json idea file.
   • The ideation script performs Semantic Scholar searches through ai_scientist/tools/semantic_scholar.py; for paper/literature queries, use the s2-api skill and load S2_API_KEY from /home/qiao/dockerai/AI-Scientist-v2/.env. Report rate-limit/key problems clearly.

3. Run a BFTS experiment

   • Use launch_scientist_bfts.py --load_ideas <ideas.json> --idea_idx <n> --model <model>.
   • Add --skip_writeup for faster debugging; add/omit --skip_review based on whether the user wants auto-review.
   • Use --load_code only when a same-stem .py file should seed the experiment.

4. Inspect results

   • Runs create timestamped directories under experiments/.
   • Look for logs/0-run/unified_tree_viz.html, experiment_results/, PDFs, review_text.txt, review_img_cap_ref.json, and token_tracker*.json.
   • Report exact paths and last meaningful log lines when diagnosing failures.

5. Use the web console

   • Launch with bash run_web.sh --port 8501 --host 127.0.0.1.
   • The Streamlit console manages ideas, background jobs, BFTS config, papers/reviews, API registrations, verification keys, and API settings.
   • Bind to 0.0.0.0 only behind HTTPS/auth; set AI_SCIENTIST_APP_PASSWORD and/or access-key gating for public/internal testing.

Safety and operations

• This project executes LLM-written Python. Confirm before running jobs that can spawn long GPU/LLM workloads.
• Never print full API keys, API_ACCESS_KEY_FERNET, encrypted tokens, or full verification-key pools. Distribute only plaintext ASIC-TST-... verification keys when explicitly requested.
• Do not kill unrelated host processes. The launcher only terminates child processes spawned by the current run; keep that behavior.
• If Semantic Scholar returns 401/429/TLS/network errors, explain the cause and suggest checking S2_API_KEY, rate limits, CA/proxy/VPN, or skipping citation-heavy stages.

Common user intents

• “Generate ideas for X” → create/check a Markdown topic file, run ideation, inspect the generated JSON.
• “Run AI Scientist on this idea” → confirm cost/model/runtime, adjust bfts_config.yaml if needed, run launcher, stream logs.
• “Open the cloud UI” → run run_web.sh, then share localhost URL.
• “Why did my run fail?” → inspect _web_jobs/*.log or experiment logs/, dependencies, API auth, CUDA/GPU state, and writeup/review artifacts.
• “Paper query / citations” → invoke the s2-api skill and query Semantic Scholar with S2_API_KEY loaded from /home/qiao/dockerai/AI-Scientist-v2/.env; fall back to other paper-search skills only if S2 is unavailable or insufficient.