extract

name	extract
type	skill
category	instruction
description	General extraction/ingestion skill that routes to specific workflows based on input type. Extracts structured information from documents, emails, reviews, feedback, and other sources.
triggers	["extract information","extract from document","ingest","extract decisions","extract training data","convert document","DOCX/PDF/XLSX conversion","/convert-to-md"]
modifies_files	true
needs_task	true
mode	execution
domain	["operations"]
allowed-tools	Read,Grep,Glob,Edit,Write,Skill,Bash
version	1.0.0
permalink	skills-extract

Extraction & Ingestion Skill

Taxonomy note: This skill provides domain expertise (HOW) for extracting structured information from documents and sources. See [[aops-core/skills/remember/references/TAXONOMY.md]] for the skill/workflow distinction.

General-purpose extraction skill that intelligently routes to specialized workflows based on input type. Extracts structured information from various sources and stores it appropriately (public framework vs. private data).

Framework Context

Universal axioms apply (enforced by rbg). P#52 (Read-Then-Write Memory) is especially relevant.

Search Before Creating

MANDATORY before creating any new extracted content: search PKB for existing knowledge on the same subject.

mcp__pkb__search(query="[topic/person/document subject]")

• If a match exists → augment it rather than creating a duplicate
• If people are mentioned → retrieve existing relationship context to enrich extraction
• If no match → proceed with creation

This prevents duplicate memories and grounds extraction in accumulated knowledge. See [[remember]] skill's "search first" step as the model.

Purpose

Provide a unified entry point for all extraction tasks:

• Training data extraction from feedback documents
• Archive information extraction (emails, correspondence)
• Decision extraction from task queues
• Knowledge extraction from documents
• Review pair extraction for LLM training

Workflow Routing

When invoked, analyze the input and route to the appropriate workflow:

1. Training Data Extraction

Signals:

• Input contains review feedback + source document
• User mentions "training", "extract patterns", "learning", "dataset"
• Documents contain tracked changes, comments, or annotations
• Goal is to build LLM training data

Route to: workflows/training-data.md

Storage:

• Sensitive data (actual review content, source documents): $ACA_DATA/processed/review_training/
• Generalized patterns (depersonalized principles): Framework docs or peer-review skill

2. Archive Information Extraction

Signals:

• Input is email archive, correspondence, receipts