Automatic Extraction

Knowledge can scan your existing sources and extract implicit rules, decisions, and constraints automatically. Nothing is published without human review.

Why Automatic Extraction?

Most teams already have rules - they're just not structured. They live in READMEs, architecture decisions, runbooks, code comments, Slack threads, and wiki pages.

Manually entering each rule into a registry takes time most teams don't have. Automatic extraction solves the cold-start problem: point it at your sources and get a populated registry in minutes.

And as your documentation evolves, you can re-extract regularly to surface new implicit rules - your registry stays current without manual maintenance.

How It Works

1. Point at your sources

> "Extract rules from ./docs and ./CLAUDE.md for the Engineering scope"

Target specific directories and file types:

> "Extract rules from ./docs (*.md), ./src (README.md), and ./CLAUDE.md for Engineering"

2. Each chunk is analyzed

Files are split into contextual chunks (~1500 characters with 10% overlap). Each chunk is analyzed with a structured extraction prompt that identifies:

• Invariant candidates : absolute constraints ("All API endpoints must require authentication")
• Rule candidates : active directives, mandatory or advisory ("Use conventional commits")
• Decision candidates : historical choices with context ("We chose PostgreSQL for transactional data")

Each extraction includes:

• Confidence score (0.0 - 1.0, minimum 0.6 to be kept)
• Source excerpt that motivated the extraction
• Suggested tags and namespace
• An explanation of why this was identified

3. Deduplication filters noise

Every extraction is compared against existing entries in the target scope using semantic similarity:

4. Human review

Nothing is published without validation. Every extraction becomes a draft visible in the dashboard, showing:

• The proposed entry with its detected type (Invariant, Rule, or Decision) and confidence score
• The source excerpt that motivated the extraction
• Detected relations to existing entries - duplicates, replacements, tensions
• An explanation of why this was identified

Three actions:

• Approve : publishes the draft as a real Knowledge entry with source: auto_extracted attribution
• Reject : discards the draft
• Edit : modify the content, tags, or namespace before approving

Source Types

Git Repositories

Ask your AI agent to scan a repository. It reads local files, chunks them, and sends them to Knowledge for analysis. Implicit rules and constraints surface - even ones that are not documented anywhere.

> "Scan /path/to/repo for .ts, .py, .yaml, and .md files in the Engineering scope"

Specific Documents

> "Extract rules from runbook.md and architecture.md for Engineering"

Ingestion API

For sources that don't live on disk, push documents directly via the API:

curl -X POST https://api.asplenz.com/knowledge/v1/extract/stream \
  -H "Authorization: Bearer kn_..." \
  -H "Content-Type: application/json" \
  -d '{
    "scope_id": "scp-...",
    "documents": [
      {
        "content": "All deployments must go through staging first.",
        "metadata": {"author": "ops-team", "source": "runbook-v3"}
      }
    ],
    "auto_run": true
  }'

Additional Connectors

Slack, Teams, Notion, Confluence, and Excel connectors are available on Team and Scale plans.

AI Configuration

Extraction requires AI access. Two options:

Organizations with strict data residency or Zero Data Retention requirements should use their own API key.

Permissions

Configuration

Best Practices

Start broad, then refine. Run extraction on your entire docs/ directory first. Review the results, then narrow your patterns to the most productive sources.

Re-extract regularly. Run extraction quarterly, after a major rewrite, or whenever new docs appear. Smart deduplication ensures your registry won't be polluted with duplicates.

Review in batches. Review all pending drafts for a run in a single session - reject the noise, approve the good ones.

Use tags consistently. The extraction suggests tags, but review them for consistency. A clean tagging system makes the registry more searchable.

Learn More

• Getting Started →
• How Knowledge Works →
• API Reference →