Extract

extract

LLM-guided extraction or transformation from a single Note. Output is grounded entirely in the input — no new information introduced.

Input

• target: Single Note (variable, ID, or name) — NOT Collections
• instruction: What to extract or how to transform (required)
• target_tokens: Optional output length in tokens
• out: Variable name for resulting Note

Planning Notes

Prefer Python for small, predictable structure (YAML frontmatter, key:value blocks). Use extract when layout is unknown or you need semantic judgment.

Output length is unreliable. Word-count instructions ("output a 1-3 word phrase") are frequently ignored — expect verbose output. When output will be used as a filename, directory name, or identifier, always truncate and sanitize in Python:

tokens = re.split(r'[\s_]+', raw)[:3]  # split on spaces AND underscores
name = re.sub(r'[^a-z0-9_]', '', '_'.join(tokens).lower())

Do NOT use when:

• Bibliography/references → extract-references (GROBID, structured, deterministic)
• Multiple documents → synthesize
• New content from scratch → generate-note
• Filtering Collection items → filter-structured or filter-semantic
• Structured metadata fields → project or pluck

Prefer over synthesize when there is only ONE source document.

Anti-Patterns

• ❌ extract(target=$collection) — Use map(extract) for Collections.
• ❌ extract(target=$paper, instruction="extract citations") — Use extract-references.
• ❌ Raw extract output as directory/file name — Always truncate + sanitize in Python first.

Examples

{"type":"extract","target":"$paper","instruction":"Extract the key architectural innovation as one sentence.","out":"$innovation"}
{"type":"extract","target":"$abstract","instruction":"Compress to 2-3 sentences retaining methodology and results.","out":"$compressed"}
{"type":"map","target":"$papers","operation":"extract","instruction":"State the main contribution in one sentence.","out":"$contributions"}