discovery-extract（发现摘录）

Discovery Engine — Paper Extraction Skill

Why This Exists

A nanofiltration paper provides selective passage below a size threshold. A drug delivery paper requires selective transport to a target. Neither team knows the other exists — they publish in different journals, use different vocabulary, and will never cite each other.

This skill extracts structured provides/requires relationships from scientific papers, building a knowledge graph that surfaces these hidden cross-domain connections. Each paper you extract adds a node. The graph finds the bridges.

When to Use

Core Concept

Every paper is decomposed into:

• Part A (Facts): entities, properties, relations — what the paper reports
• Part B (Cross-domain): core friction, mechanism, bridge tags, provides/requires interface, unsolved tensions — what connects it to other fields

The cross_domain section is where discovery happens. The provides and requires fields use abstract functional language (not domain jargon) so a materials science paper can match a biology paper.

You Are the Extractor

No external API keys or LLM calls needed — you read the paper text and produce the structured JSON yourself. The bundled prompt (references/prompt.txt) is your extraction specification.

How It Works

1. Run python scripts/extract.py discover to find new papers with abstracts
2. Read references/prompt.txt — the full extraction format specification
3. For each paper: read its abstract and produce the extraction JSON following the prompt
4. Save each result via python scripts/extract.py save
5. Optionally submit results as a PR via gh

Step 1: Discover Papers

python scripts/extract.py discover --count 5

This outputs a JSON array of papers (id, source, title, abstract) to stdout. Already-processed papers are automatically excluded.

To target a specific source:

python scripts/extract.py discover --source arxiv --count 5
python scripts/extract.py discover --source pmc --count 5

Step 2: Read the Extraction Prompt

Read references/prompt.txt to understand the output format. It specifies:

• Part A (Facts): entities, properties, relations
• Part B (Cross-domain): core_friction, mechanism, bridge_tags, provides/requires interface, unsolved_tensions

The prompt contains detailed rules, examples, and a self-check procedure.

Step 3: Extract

For each paper from Step 1, produce a JSON object following the schema in references/prompt.txt. The paper's abstract is your input text.

Write the JSON to a temporary file (e.g., /tmp/result.json or any local path).

Key requirements:

• Output ONLY valid JSON (no markdown wrapping, no commentary)
• The top-level key must be paper_analysis (not analysis)
• unsolved_tensions entries must be objects with {tension, constraint_class, why_it_matters, source_quote}
• provides entries must be objects with {operation, description, performance, conditions}
• requires entries must be objects with {operation, description, reason}
• bridge_tags must be abstract functional descriptors, not domain nouns
• The cross_domain section is where discovery happens — invest effort here

Step 4: Save Results

python scripts/extract.py save /tmp/result.json \
  --paper-id "arxiv:2401.00001" \
  --source arxiv \
  --title "Paper Title Here"

The save command normalizes format issues, validates, adds metadata, and saves to ~/.discovery/data/batch/. It will report any validation warnings.

Step 5: Validate (optional)

python scripts/extract.py validate ~/.discovery/data/batch/

Step 6: Submit Results (optional)

After extracting a batch, submit results as a PR:

# Fork (first time only)
gh repo fork pcdeni/discovery-engine --clone=false

# Clone your fork
gh repo clone pcdeni/discovery-engine discovery-engine-submit
cd discovery-engine-submit

# Create branch and copy results
BRANCH="contrib/$(gh api user --jq .login)/$(date +%Y%m%d-%H%M%S)"
git checkout -b "$BRANCH"
cp ~/.discovery/data/batch/*.json submissions/
git add submissions/
git commit -m "Add extraction results"
git push -u origin "$BRANCH"

# Create PR
gh pr create --title "extraction: $(ls submissions/*.json | wc -l) papers" \
  --body "Extraction results from discovery-extract skill" \
  --repo pcdeni/discovery-engine

GitHub Actions CI validates submissions and auto-merges passing PRs.

Bundled Files