Sciatlas Idea Generate

SciAtlas Idea Generate

Use this skill to generate research ideas from SciAtlas KG evidence. Run only the repository workflow python -m sciatlas_idea_gen.main, which retrieves anchor papers, builds a research graph, searches same-field and cross-domain inspirations, generates ideas, and runs novelty checks.

Operating Contract

• Own the end-to-end novice flow: install or locate the workflow, guide registration, configure .env or shell variables, run retrieval/generation, inspect artifacts, and synthesize final idea seeds.
• Ask the user only for human-only values: missing topic/domain, email, verification code, SciAtlas token, LLM key/base/model, S2 key, local KG credentials when explicitly using local KG, or one necessary scope clarification.
• Do not ask the user to run shell commands when tool access is available.
• Never disclose the full API token.
• Use sciatlas_idea_gen as the only idea-generation workflow.
• If sciatlas_idea_gen is unavailable or required LLM/KG setup is missing, report the blocker and fix setup if possible; do not substitute another route.
• Save and inspect artifacts before writing the final answer.

KG Linkage

The workflow reaches the KG through sciatlas_idea_gen.clients.SciAtlasClient:

• Hosted mode: SCIATLAS_USE_LOCAL_KG=0; the adapter shells into this repository's run_sciatlas.py for KG-backed paper retrieval, then normalizes data.result.ranking.papers or data.result.sources.kg.papers to papers.
• Local KG mode: SCIATLAS_USE_LOCAL_KG=1; the adapter runs references/search/run_search.py with --disable-s2, --disable-llm-ranking, --kg-top-k, and --final-top-k, then normalizes the local KG response to papers.
• Keep hosted mode as the default unless the user says they have Neo4j and local KG models configured.

Zero-Start Bootstrap

1. Check for the repository workflow:

python -m sciatlas_idea_gen.main -h

2. If missing, clone the full SciAtlas repository, change into it, then run python -m pip install -e ./sciatlas and python -m pip install -r requirements-workflows.txt. Do not use the GitHub #subdirectory=sciatlas package-only installation: it does not include sciatlas_idea_gen.
3. Check current environment and .env for hosted KG and LLM settings before asking the user.
4. If SCIATLAS_API_KEY is missing for hosted mode, guide the user to http://sciatlas.openkg.cn/register; ask for email, verification code, and returned sciatlas_xxx token only when needed.
5. Configure hosted KG yourself:

$env:SCIATLAS_API_BASE_URL = "http://sciatlas.openkg.cn"
$env:SCIATLAS_API_KEY = "<token>"
$env:SCIATLAS_USE_LOCAL_KG = "0"

export SCIATLAS_API_BASE_URL="http://sciatlas.openkg.cn"
export SCIATLAS_API_KEY="<token>"
export SCIATLAS_USE_LOCAL_KG=0

6. Configure the LLM endpoint required by the multi-step workflow:

export LLM_API_KEY="<provider-key>"
export LLM_BASE_URL="https://your-provider.example/v1"
export LLM_MODEL="<model-name>"

Ask only for missing LLM provider values, then write them to the current shell or .env without echoing secrets.

7. Configure local KG only when requested:

export SCIATLAS_USE_LOCAL_KG=1
export SCIATLAS_LOCAL_SEARCH_ROOT="./references/search"

The local KG engine also needs its own references/search/.env with Neo4j/model settings.

Run Plan

Use the flash path first for quick, interactive idea generation:

python -m sciatlas_idea_gen.main "<research topic>" --workflow flash --domain "<optional field>"

Run the full path when the user wants a broader research graph, more seed diversity, and a more comprehensive inspiration search:

python -m sciatlas_idea_gen.main "<research topic>" --workflow full --domain "<optional field>"

For a seed PDF:

python -m sciatlas_idea_gen.main "<research topic>" --workflow full --pdf path/to/paper.pdf

Useful overrides:

• --workflow flash for a faster path: 1 seed query, a compact graph, one same-field inspiration, one cross-domain inspiration, compressed Step 5/Step 8 gate stages, and no novelty feedback retry.
• --workflow full for the broader path: multi-query seed retrieval, larger graph construction, broader inspiration retrieval, and novelty feedback retry.
• --k-step1 N for more seed papers.
• --anchor-top-k N for broader Step 1 retrieval.
• --graph-budget-min N and --graph-budget-max N for graph size.
• --seed-recent-years N to include older foundational papers.
• --resume-from-run-dir runs/<id> to continue from saved artifacts.

Artifacts To Read

Read the run directory before answering:

• summary.json: status, effective config, counts, and final ideas.
• retrieval_trace.json: KG/S2/OpenAlex retrieval events and returned papers.
• step1_seed_papers.json: anchor papers and refined query.
• step2_research_graph.json: graph nodes, edges, evidence scores, and phases.
• step3_trend.txt: trend summary from the graph.
• step5_radius_plan.json: inspiration radius plan. In flash, this is a deterministic compact plan and summary.json may mark Step 5 as compressed.
• step6_inspiration_candidates.json, step7_inspirations.json, step8_selected_inspirations.json: inspiration path.
• step9_ideas.json and step9_ideas.md: final ideas and citations.

In flash, treat status: "compressed" for Step 5 or Step 8 as normal success. Step 5 skips the LLM radius gate and uses the compact same-field/cross-domain plan. Step 8 skips the LLM selector and keeps the compact plausible inspiration set. Do not rerun full merely because these stages are compressed.

If the run is in smoke mode and fails partway, use the completed artifacts and clearly state the failed step. Smoke mode is now diagnostic; prefer --workflow flash for normal quick runs.

Deliverable

Return 3-8 idea seeds when available. For each idea:

• Name.
• One-sentence hypothesis.
• KG evidence trail from seed papers, graph papers, and inspirations.
• Why it is nontrivial.
• Closest-prior-art or novelty risk.
• Minimal validation experiment.
• Next SciAtlas query or pipeline setting to test novelty.

Keep speculative claims clearly labeled and cite artifacts/paper titles from the run.