Flip RAG SDK 🚀

Fully Automated Retrieval-Augmented Generation

Flip is the simplest yet most powerful RAG SDK. Initialize with a directory and start querying - that's it!

✨ Features

• 🎯 Ultra-Simple API: Just Flip(directory="./docs") to get started
• 🔄 Fully Automated: Handles everything from document loading to response generation
• 🤖 Multiple LLM Providers: OpenAI, Azure OpenAI, Anthropic, Google, HuggingFace, Meta, Ollama
• 📊 Smart Chunking: Token, sentence, semantic, and recursive strategies
• 🔍 Hybrid Search: Vector + keyword search for better accuracy
• 💾 9 Vector Databases: Pinecone, Qdrant, Weaviate, Milvus, FAISS, Pgvector, Redis, Elasticsearch, MongoDB
• 📝 Rich File Support: PDF, DOCX, TXT, MD, JSON, CSV, HTML, code files
• 🎨 Extensible: Easy to add custom providers and strategies

🚀 Advanced Features

Flip includes production-ready advanced features:

• 🔀 Hybrid Search: Combines vector similarity + BM25 keyword search with reciprocal rank fusion
• 📊 Re-ranking: Cross-encoder models for improved retrieval accuracy
• 🧠 Query Enhancement: Automatic query classification, rewriting, decomposition, and HyDE
• 💾 Intelligent Caching: LRU caching for embeddings and queries with disk persistence
• ⚡ Pipeline Orchestration: Seamless integration of all components
• 🔄 Incremental Updates: Efficient document change tracking
• 📈 Performance Monitoring: Query logging and analytics
• 📊 Built-in Evaluation: Precision, recall, F1, MRR, NDCG metrics

See ADVANCED_FEATURES.md for detailed documentation.

🚀 Quick Start

Installation

pip install flip-rag

Basic Usage

from flip import Flip

# Initialize with a directory - that's it!
flip = Flip(directory="./docs")

# Query your documents
response = flip.query("What is the main topic?")

print(response.answer)
print(response.citations)

Choose Your LLM Provider

# Use OpenAI (default)
flip = Flip(directory="./docs", llm_provider="openai")

# Use Azure OpenAI
flip = Flip(directory="./docs", llm_provider="azure-openai")

# Use Anthropic Claude
flip = Flip(directory="./docs", llm_provider="anthropic")

# Use Google Gemini
flip = Flip(directory="./docs", llm_provider="google")

# Use HuggingFace
flip = Flip(directory="./docs", llm_provider="huggingface")

# Use local Ollama
flip = Flip(directory="./docs", llm_provider="ollama")

Advanced Configuration

from flip import Flip, FlipConfig

# Custom configuration
config = FlipConfig(
    llm_provider="openai",
    llm_model="gpt-4-turbo-preview",
    embedding_provider="openai",
    chunking_strategy="semantic",
    chunk_size=512,
    retrieval_top_k=5,
    use_hybrid_search=True,
)

flip = Flip(directory="./docs", config=config)

Azure OpenAI Configuration

from flip import Flip, FlipConfig

config = FlipConfig(
    llm_provider="azure-openai",
    llm_model="gpt-4",  # Your deployment name
    
    embedding_provider="azure-openai",
    embedding_model="text-embedding-ada-002",  # Your deployment name
    
    azure_openai_api_key="your-key",
    azure_openai_endpoint="https://your-resource.openai.azure.com/",
)

flip = Flip(directory="./docs", config=config)
response = flip.query("What is AI?")

📚 Supported LLM Providers

Provider	Models	API Key Required
OpenAI	GPT-4, GPT-3.5	Yes (OPENAI_API_KEY)
Azure OpenAI	GPT-4, GPT-3.5	Yes (AZURE_OPENAI_API_KEY)
Anthropic	Claude 3 (Opus, Sonnet, Haiku)	Yes (ANTHROPIC_API_KEY)
Google	Gemini Pro	Yes (GOOGLE_API_KEY)
HuggingFace	Llama 2, Mistral, etc.	Yes (HUGGINGFACE_API_KEY)
Meta	Llama models via HuggingFace	Yes (HUGGINGFACE_API_KEY)
Ollama	Any local model	No (runs locally)

🔗 Supported Embedding Providers

Provider	Models	API Key Required
OpenAI	text-embedding-3-small/large	Yes
Azure OpenAI	text-embedding-ada-002	Yes
Google	models/embedding-001	Yes
HuggingFace	Various models	Yes
Sentence Transformers	all-MiniLM-L6-v2, etc.	No (local)

📄 Supported File Formats

• Documents: PDF, DOCX, TXT, MD
• Data: JSON, CSV
• Web: HTML, HTM
• Code: Python, JavaScript, TypeScript, Java, C++, Go, Rust, Ruby

🔧 API Reference

Flip Class

flip = Flip(
    directory: Optional[str] = None,
    config: Optional[FlipConfig] = None,
    llm_provider: Optional[str] = None,
    llm_model: Optional[str] = None,
    **kwargs
)

Methods

• query(question: str) -> FlipResponse: Query your documents
• index_directory(directory: str): Index a directory
• add_documents(file_paths: List[str]): Add specific files
• refresh_index(): Incrementally update index with changed documents
• evaluate(query, relevant_ids, k=5): Evaluate RAG performance
• get_monitoring_stats(): Get performance statistics
• clear(): Clear all indexed documents
• get_stats(): Get indexing statistics

FlipResponse

@dataclass
class FlipResponse:
    answer: str                    # Generated answer
    citations: List[Dict]          # Source citations
    context_chunks: List[str]      # Retrieved chunks
    metadata: Dict[str, Any]       # Additional metadata

FlipConfig

config = FlipConfig(
    # LLM Configuration
    llm_provider="openai",         # openai, azure-openai, anthropic, google, huggingface, ollama
    llm_model=None,                # Auto-selected if None
    llm_temperature=0.7,
    llm_max_tokens=1024,
    
    # Embedding Configuration
    embedding_provider="openai",   # openai, azure-openai, google, sentence-transformers, huggingface
    embedding_model=None,          # Auto-selected if None
    
    # Azure OpenAI (if using azure-openai provider)
    azure_openai_api_key=None,
    azure_openai_endpoint=None,
    azure_openai_api_version="2024-02-15-preview",
    
    # Vector Store
    vector_store="chroma",         # chroma (more coming soon)
    vector_store_path=None,        # Auto: ./flip_data
    
    # Chunking
    chunking_strategy="semantic",  # token, sentence, semantic, recursive
    chunk_size=512,
    chunk_overlap=50,
    
    # Retrieval
    retrieval_top_k=5,
    use_hybrid_search=True,
    use_reranking=True,
    
    # Cache
    enable_cache=True,
    cache_dir=None,                # Auto: ./flip_cache
)

🌟 Examples

Example 1: Simple Q&A

from flip import Flip

flip = Flip(directory="./company_docs")
response = flip.query("What is our refund policy?")

print(f"Answer: {response.answer}")
print(f"Sources: {len(response.citations)} documents")

Example 2: Multiple Providers

from flip import Flip

# Try different providers
providers = ["openai", "azure-openai", "anthropic", "google"]

for provider in providers:
    flip = Flip(directory="./docs", llm_provider=provider)
    response = flip.query("Summarize the main points")
    print(f"{provider}: {response.answer[:100]}...")

Example 3: Incremental Updates

from flip import Flip

flip = Flip(directory="./docs")

# Later, refresh with only changed documents
stats = flip.refresh_index()

print(f"Added: {stats['added']}")
print(f"Updated: {stats['updated']}")
print(f"Deleted: {stats['deleted']}")

Example 4: Evaluation

from flip import Flip

flip = Flip(directory="./docs")

# Evaluate a query
result = flip.evaluate(
    query="What is AI?",
    relevant_doc_ids=["chunk_1", "chunk_2"],  # Ground truth
    k=5
)

print(f"Precision: {result.retrieval_precision:.3f}")
print(f"Overall Score: {result.overall_score:.3f}")

🔑 Environment Variables

Set API keys as environment variables:

export OPENAI_API_KEY="your-key-here"
export AZURE_OPENAI_API_KEY="your-azure-key-here"
export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
export ANTHROPIC_API_KEY="your-key-here"
export GOOGLE_API_KEY="your-key-here"
export HUGGINGFACE_API_KEY="your-key-here"

Or use a .env file:

OPENAI_API_KEY=your-key-here
AZURE_OPENAI_API_KEY=your-azure-key-here
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
ANTHROPIC_API_KEY=your-key-here
GOOGLE_API_KEY=your-key-here
HUGGINGFACE_API_KEY=your-key-here

🗄️ Multi-Database Support

Flip supports 9 major vector and NoSQL databases, giving you complete flexibility to choose the perfect storage backend for your use case:

Vector Databases

Pinecone (Cloud-Native)

config = FlipConfig(
    vector_store="pinecone",
    pinecone_api_key="your-key",
    pinecone_environment="gcp-starter",
    pinecone_index_name="flip-index"
)
flip = Flip(directory="./docs", config=config)

Qdrant (Local/Cloud)

config = FlipConfig(
    vector_store="qdrant",
    qdrant_url="http://localhost:6333",  # or cloud URL
    qdrant_api_key="your-key"  # for cloud
)
flip = Flip(directory="./docs", config=config)

Weaviate (GraphQL)

config = FlipConfig(
    vector_store="weaviate",
    weaviate_url="http://localhost:8080",
    weaviate_api_key="your-key"  # for cloud
)
flip = Flip(directory="./docs", config=config)

Milvus (Enterprise)

config = FlipConfig(
    vector_store="milvus",
    milvus_host="localhost",
    milvus_port=19530
)
flip = Flip(directory="./docs", config=config)

FAISS (Local, Ultra-Fast)

config = FlipConfig(
    vector_store="faiss",
    faiss_index_type="HNSW",  # or IVF, Flat
    persist_directory="./faiss_data"
)
flip = Flip(directory="./docs", config=config)

Pgvector (PostgreSQL)

config = FlipConfig(
    vector_store="pgvector",
    pgvector_host="localhost",
    pgvector_database="flip_db",
    pgvector_user="postgres",
    pgvector_password="password"
)
flip = Flip(directory="./docs", config=config)

NoSQL/Hybrid Databases

Redis (In-Memory + Cache)

config = FlipConfig(
    vector_store="redis",
    redis_host="localhost",
    redis_port=6379
)
flip = Flip(directory="./docs", config=config)

Elasticsearch (Search Engine)

config = FlipConfig(
    vector_store="elasticsearch",
    elasticsearch_url="http://localhost:9200",
    elasticsearch_api_key="your-key"
)
flip = Flip(directory="./docs", config=config)

MongoDB (Document Store)

config = FlipConfig(
    vector_store="mongodb",
    mongodb_uri="mongodb://localhost:27017/",
    mongodb_database="flip_db"
)
flip = Flip(directory="./docs", config=config)

Database Comparison

Database	Type	Best For	Deployment
Pinecone	Vector	Production, cloud-native	Cloud
Qdrant	Vector	Flexibility, snapshots	Local/Cloud
Weaviate	Vector	GraphQL, schema management	Local/Cloud
Milvus	Vector	Enterprise, scalability	Local/Cloud
FAISS	Vector	Local, speed, GPU support	Local
Pgvector	Vector	PostgreSQL users, ACID	Local/Cloud
Redis	Hybrid	Caching, real-time	Local/Cloud
Elasticsearch	Hybrid	Full-text + vectors	Local/Cloud
MongoDB	NoSQL	Metadata-rich storage	Local/Cloud

Key Features Across All Databases

• ✅ Unified API: Same interface for all databases
• ✅ Lazy Loading: Optional dependencies (install only what you need)
• ✅ Health Checks: Monitor database status
• ✅ Batch Operations: Efficient bulk inserts
• ✅ Metadata Filtering: Filter by custom metadata
• ✅ Statistics: Track vector counts and performance

See individual database examples in the examples/ directory.

🎯 Design Philosophy

Flip is designed around three core principles:

1. Simplicity First: The API should be so simple that you can get started in 2 lines of code
2. Automation: Handle all the complexity internally - chunking, embedding, retrieval, generation
3. Flexibility: Support multiple providers and allow customization when needed

🛠️ Advanced Features

Custom Providers

You can register custom LLM or embedding providers:

from flip.generation.factory import LLMFactory
from flip.generation.base import BaseLLM

class MyCustomLLM(BaseLLM):
    # Implement required methods
    pass

LLMFactory.register_provider("mycustom", MyCustomLLM)

# Use it
flip = Flip(directory="./docs", llm_provider="mycustom")

📊 Performance

Flip is optimized for performance:

• Batch Processing: Embeddings generated in batches
• Caching: Automatic caching of embeddings and queries
• Efficient Storage: Support for 9 optimized vector databases
• Incremental Updates: Only re-index changed documents
• Performance Testing: Built-in benchmarking framework

🧪 Testing

Flip includes comprehensive testing infrastructure:

# Use test fixtures for any database
from tests.fixtures import all_vector_stores

def test_my_feature(all_vector_stores):
    # Test runs against all available databases
    pass

# Benchmark performance
from tests.performance import PerformanceTester

tester = PerformanceTester(vector_store)
tester.run_full_benchmark()
tester.print_results()

🤝 Contributing

Contributions are welcome! Please see our contributing guidelines.

📝 License

MIT License - see LICENSE file for details

🙏 Acknowledgments

Flip builds on the shoulders of giants:

• OpenAI, Azure OpenAI, Anthropic, Google for LLM APIs
• Pinecone, Qdrant, Weaviate, Milvus, FAISS, Pgvector, Redis, Elasticsearch, MongoDB for vector storage
• Sentence Transformers for local embeddings
• And many other open-source projects

📞 Support

• Documentation: https://flip-rag.readthedocs.io
• Issues: GitHub Issues
• Discussions: GitHub Discussions

---

Made with ❤️ by the Flip Team