bdistill-knowledge-extraction

استخراج دانش ساختاریافته از مدل‌ها در نشست یا با Ollama

این اسکیل دانش ساختاریافتهٔ یک حوزه را در همان نشست از مدل AI یا از مدل‌های متن‌باز محلی با Ollama استخراج می‌کند. کلید API لازم نیست.

چه چیزی می‌گیرید

  • نصب
  • استخراج در نشست
  • جست‌وجو و خروجی

برای چه کاری خوب است

  • ساخت پایگاه دانش
  • مدل‌های محلی

برای چه کاری نیست

  • دانش حساس محرمانه

دربارهٔ این اسکیل

Knowledge Extraction

Extract structured, quality-scored domain knowledge from any AI model — in-session from closed models (no API key) or locally from open-source models via Ollama.

Overview

bdistill turns your AI subscription sessions into a compounding knowledge base. The agent answers targeted domain questions, bdistill structures and quality-scores the responses, and the output accumulates into a searchable, exportable reference dataset.

Adversarial mode challenges the agent's claims — forcing evidence, corrections, and acknowledged limitations — producing model-generated candidates for independent source review; adversarial self-correction is not external validation.

When to Use This Skill

  • Use when you need structured reference data on any domain (medical, legal, finance, cybersecurity)
  • Use when building lookup tables, Q&A datasets, or research corpora
  • Use when generating training data for traditional ML models (regression, classification — NOT competing LLMs)
  • Use when you want cross-model comparison on domain knowledge

How It Works

Step 1: Install

pip install bdistill
claude mcp add bdistill -- bdistill-mcp   # Claude Code

Step 2: Extract knowledge in-session

/distill medical cardiology                    # Preset domain
/distill --custom kubernetes docker helm       # Custom terms
/distill --adversarial medical                 # With adversarial validation

Step 3: Search, export, compound

bdistill kb list                               # Show all domains
bdistill kb search "atrial fibrillation"       # Keyword search
bdistill kb export -d medical -f csv           # Export as spreadsheet
bdistill kb export -d medical -f markdown      # Readable knowledge document

Output Format

Structured reference JSONL — not training data:

{
  "question": "What causes myocardial infarction?",
  "answer": "Myocardial infarction results from acute coronary artery occlusion...",
  "domain": "medical",
  "category": "cardiology",
  "tags": ["mechanistic", "evidence-based"],
  "quality_score": 0.73,
  "confidence": 0.73,
  "validated": false,
  "source_model": "<actual host model recorded at run time>"
}

The numeric scores are illustrative, uncalibrated metadata; independent source verification is absent in this example.

Tabular ML Data Generation

Generate structured training data for traditional ML models:

/schema sepsis | hr:float, bp:float, temp:float, wbc:float | risk:category[low,moderate,high,critical]

Exports as CSV ready for pandas/sklearn. Each row tracks source_model for cross-model analysis.

Local Model Extraction (Ollama)

For open-source models running locally:

# Install Ollama from https://ollama.com
ollama serve
ollama pull qwen3:4b

bdistill extract --domain medical --model qwen3:4b

Security & Safety Notes

  • In-session extraction uses your existing subscription — no additional API keys
  • Local extraction runs entirely on your machine via Ollama
  • In-session extraction uses the host provider and its data policy; it is not offline. A genuinely local Ollama endpoint keeps inference local, but downloading models/installing packages uses the network. Verify endpoints and approved data processing before use
  • Output is reference data, not LLM training format

Related Skills

  • @bdistill-behavioral-xray - X-ray a model's behavioral patterns

Limitations

  • Generated medical, legal, financial or security content needs dated independent authoritative evidence and domain review; quality_score/confidence are not clinical/legal accuracy or calibrated probabilities. Do not put personal records into generated corpora or use synthetic labels as real patient outcomes.
  • Installing MCP or downloading models is separate from extraction; use a reviewed environment and respect host/configuration authorization.
  • Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
  • Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.

دانلود ZIP