LLMRAG-NER

中文版

Overview

LLMRAG-NER is a Retrieval-Augmented Generation (RAG) based Named Entity Recognition system that leverages Large Language Models (LLM) for entity extraction.

Features

• RAG-based NER: Uses vector similarity search to retrieve relevant examples
• Multi-entity Support: Recognizes PER (Person), LOC (Location), ORG (Organization)
• Flexible LLM Integration: Supports multiple LLM backends via factory pattern
• Voting Mechanism: Ensemble voting for improved accuracy

Project Structure

File	Description
VectorDB.py	Vector database for similarity search
preprocess.py	BIO to JSON data preprocessing
main.py	Main RAG pipeline
llm_client.py	LLM client interface
evaluate.py	Evaluation metrics
vote.py	Voting mechanism
conf/	Configuration files
data/	Training and dev datasets
result/	Output results
utils/	Utilities

Quick Start

# Process BIO format data to JSON
python preprocess.py

# Run RAG-NER pipeline
python main.py

Workflow

1. Preprocess: Convert BIO format to JSON
2. Vectorize: Sentence splitting with spaCy, embedding with HuggingFace
3. Retrieve: Search top-k similar examples by vector similarity
4. Augment: Include retrieved examples in prompt
5. Extract: Call LLM for entity recognition
6. Output: Save recognition results

Dependencies

• Python 3.x
• spaCy (zh_core_web_sm)
• LangChain
• ChromaDB
• HuggingFace sentence-transformers