AutoMetaRAG

Automatic Metadata-based Retrieval Augmented Generation framework for optimized document retrieval using hybrid search (semantic + metadata filtering).

Features

• LLM-Powered Metadata Extraction: Automatically extracts structured metadata from documents
• Hybrid Search: Combines semantic similarity search with metadata filtering
• Qdrant Vector Database: Efficient vector storage and retrieval
• Configurable Pipeline: Easy configuration through environment variables

Installation

1. Install dependencies:

pip install -r requirements.txt

2. Set up environment variables using .env file:

# Copy the example file
cp env.example .env

# Edit .env with your actual credentials

Your .env file should contain:

OPENAI_API_KEY=sk-your-openai-api-key
QDRANT_URL=https://your-cluster.qdrant.io
QDRANT_API_KEY=your-qdrant-api-key

# Optional
QDRANT_COLLECTION=AutoMetaRAG  # default: AutoMetaRAG
DATA_DIR=./data                # default: ./data

3. Configure metadata settings in config.ini:

[Metadata]
probable_questions = "Your example questions"
document_info = "Description of your dataset"

Usage

1. Ingestion Pipeline

Process documents, extract metadata, and create vector database:

python -m AutoMetaRAG --mode ingest

Optional arguments:

python -m AutoMetaRAG --mode ingest --config custom.ini --data-dir ./my_docs

Available CLI Arguments:

Argument	Mode	Type	Default	Description
--mode	Both	Required	-	ingest or query
--config	Both	Optional	config.ini	Path to config file
--data-dir	ingest	Optional	./data	Data directory path
--query	query	Optional	-	Query string (if not provided, enters interactive mode)
--score-threshold	query	Optional	0.3	Minimum relevance score (0.0-1.0)
--vector-name	query	Optional	""	Vector name for multi-vector search

2. Query Pipeline

Option A: Single query (direct)

python -m AutoMetaRAG --mode query --query "What is this paper about?"

With custom score threshold:

python -m AutoMetaRAG --mode query --query "Your question" --score-threshold 0.5

With multi-vector approach:

python -m AutoMetaRAG --mode query --query "Your question" --vector-name my_vector

Option B: Interactive session

python -m AutoMetaRAG --mode query

Then enter your questions interactively. Type 'exit' to quit.

Interactive with custom threshold:

python -m AutoMetaRAG --mode query --score-threshold 0.5

Configuration

Environment Variables (.env file)

AutoMetaRAG uses python-dotenv to load environment variables from a .env file.

Required variables:

• OPENAI_API_KEY: Your OpenAI API key
• QDRANT_URL: Qdrant server URL (e.g., https://your-cluster.qdrant.io)
• QDRANT_API_KEY: Qdrant API key

Optional variables:

• QDRANT_COLLECTION: Collection name (default: "AutoMetaRAG")
• DATA_DIR: Directory containing documents (default: "./data")

Setup:

1. Copy env.example to .env
2. Fill in your actual credentials
3. The .env file is automatically loaded when running AutoMetaRAG

config.ini File

The config.ini file contains metadata configuration for schema generation:

[Metadata]
probable_questions = "Question 1?", "Question 2?", "Question 3?"
document_info = "Description of your dataset"

Project Structure

.
├── AutoMetaRAG/        # Main package directory
│   ├── __init__.py
│   ├── __main__.py     # CLI entry point
│   ├── pipeline.py     # Main pipeline class
│   ├── config.py       # Configuration management
│   ├── metadata.py     # Metadata generation/extraction
│   ├── document.py     # Document processing
│   ├── indexer.py      # Qdrant indexing
│   ├── query.py        # Query engine
│   └── utils.py         # Utility functions
├── examples/           # Example scripts
│   ├── test1.py        # Query example
│   └── test2.py        # Ingestion example
├── requirements.txt    # Python dependencies
├── env.example         # Environment variables template
├── README.md           # This file
└── USAGE.md            # Detailed usage guide

How It Works

1. Environment Setup: Loads API keys and configuration from .env file using python-dotenv
2. Metadata Schema Generation: LLM analyzes your data and suggests metadata schemas
3. Document Processing: Documents are loaded and metadata is extracted using LLM
4. Vector Database Creation: Documents are embedded and stored in Qdrant with metadata
5. Query Processing: User queries are analyzed to extract metadata filters
6. Hybrid Search: Uses Qdrant's query_points method with:
   • Semantic similarity search
   • Metadata filtering
   • Score threshold filtering (minimum relevance score: 0.3)
   • Multi-vector support (optional)
7. Response Generation: LLM generates answers from retrieved context

Example

from AutoMetaRAG import AutoMetaRAGPipeline

# Initialize pipeline
pipeline = AutoMetaRAGPipeline('config.ini')

# Run ingestion
pipeline.run_indexing_pipeline()

# Query
answer = pipeline.query("What did the paper discuss about transformers?")
print(answer)