LLM Agents for Technical Metadata Extraction

Hybrid ML + LLM pipeline that automatically extracts and validates technical metadata from web sources and internal databases. Combines Random Forest & SVM classifiers with LLM agents for high-accuracy, structured output from unstructured engineering text.

---

🔍 Project Overview

A two-stage intelligence system:

1. ML Stage (fast, local): Random Forest + SVM classifiers trained on TF-IDF features to classify component type with confidence scores
2. LLM Agent Stage (deep understanding): LLM validates, corrects, and enriches metadata — filling gaps the ML model misses

Output: Structured JSON with component type, voltage, capacity, weight, certifications, manufacturer, part number, and more — with a full audit trail of corrections.

---

🏗️ Architecture

llm-agents-metadata-extraction/
├── src/
│   ├── llm_agent.py       # LLM agent + web scraper + rule-based extractor
│   ├── ml_classifier.py   # RF + SVM ensemble classifier
│   └── pipeline.py        # End-to-end orchestration
├── data/
├── configs/
├── requirements.txt
└── README.md

---

⚙️ Setup

git clone https://github.com/PRATdoppelEK/llm-agents-metadata-extraction.git
cd llm-agents-metadata-extraction
pip install -r requirements.txt

For local LLM (recommended for privacy):

# Install Ollama: https://ollama.ai
ollama pull mistral

For OpenAI API:

export OPENAI_API_KEY=sk-...

---

🚀 Quickstart

Demo mode (no LLM required — runs ML only)

python src/pipeline.py

With URLs

python src/pipeline.py \
  --urls https://www.batteryspace.com/prod-specs/... \
  --llm_url http://localhost:11434/v1/chat/completions \
  --output results/metadata.json

With raw text input

python src/pipeline.py \
  --texts "Samsung 50Ah NMC cell 3.65V UN38.3 certified" \
  --api_key $OPENAI_API_KEY \
  --llm_url https://api.openai.com/v1/chat/completions \
  --llm_model gpt-4o-mini

Batch JSON input

[
  {"id": "comp_001", "url": "https://example.com/battery-spec"},
  {"id": "comp_002", "text": "LFP cell 3.2V 100Ah automotive grade..."}
]

python src/pipeline.py --input_json batch.json --output results/batch_metadata.json

---

📊 Sample Output

{
  "item_id": "demo_cell",
  "component_type": "battery_cell",
  "material": "NMC",
  "voltage_v": 3.65,
  "capacity_ah": 50.0,
  "weight_kg": 0.9,
  "temperature_range": "-30°C to 60°C",
  "certifications": ["UN 38.3"],
  "manufacturer": "Samsung SDI",
  "part_number": "INR21700-50E",
  "ml_label": "battery_cell",
  "ml_confidence": 0.94,
  "llm_validated": true,
  "llm_corrections": {}
}

ML classifier achieves ~94% confidence on battery component classification. LLM agent corrects and enriches ~30% of ML outputs with additional structured fields.

🧠 Key Technical Highlights

• Ensemble ML: RF + SVM with TF-IDF (1–3 gram) for robust text classification
• LLM Validation: Structured JSON extraction with audit trail of corrections
• Privacy-first: Works fully offline with local LLMs (Ollama/Mistral/Llama)
• Rule-based pre-pass: Regex extraction for voltage, capacity, certifications before LLM
• Web scraping: BeautifulSoup-based scraper for product pages

---

🔧 Tech Stack

scikit-learn · LangChain-compatible API · BeautifulSoup4 · Ollama · OpenAI API · Python 3.10+

---

👤 Author

Prateek Gaur — ML Engineer | Battery & Engineering AI
LinkedIn · GitHub