Using Retrieval Augmented Generation (RAG)¶

Genie Tooling provides a flexible RAG system through the genie.rag interface, allowing you to index data from various sources and perform semantic searches.

Core RAG Operations via `genie.rag`¶

The genie.rag interface simplifies common RAG tasks:

genie.rag.index_directory(path, collection_name, ...): Indexes all supported files within a local directory.
genie.rag.index_web_page(url, collection_name, ...): Fetches content from a web URL, extracts text, and indexes it.
genie.rag.search(query, collection_name, top_k, ...): Performs a semantic search against an indexed collection.

Configuring RAG Components¶

RAG components (Document Loaders, Text Splitters, Embedding Generators, Vector Stores, Retrievers) are configured primarily using FeatureSettings within your MiddlewareConfig.

import asyncio
from pathlib import Path
from genie_tooling.config.models import MiddlewareConfig
from genie_tooling.config.features import FeatureSettings
from genie_tooling.genie import Genie

async def main():
    app_config = MiddlewareConfig(
        features=FeatureSettings(
            # RAG Embedder (e.g., for creating embeddings of your documents)
            rag_embedder="sentence_transformer", # Alias for "sentence_transformer_embedder_v1"
            # rag_embedder_st_model_name="all-MiniLM-L6-v2", # Default ST model

            # RAG Vector Store (e.g., for storing and searching embeddings)
            rag_vector_store="faiss", # Alias for "faiss_vector_store_v1" (in-memory)
            # Or for persistent ChromaDB:
            # rag_vector_store="chroma",
            # rag_vector_store_chroma_path="./my_rag_db",
            # rag_vector_store_chroma_collection_name="my_documents",
            # Or for Qdrant:
            # rag_vector_store="qdrant", # Example using Qdrant
            # rag_vector_store_qdrant_url="http://localhost:6333", # Or path for local Qdrant
            # rag_vector_store_qdrant_collection_name="my_qdrant_docs",
            # rag_vector_store_qdrant_embedding_dim=384, # Set to your embedder's dimension (e.g., 384 for all-MiniLM-L6-v2)

            # Default RAG Loader (used by index_directory if not specified)
            rag_loader="file_system", # Alias for "file_system_loader_v1" (default for index_directory)
            # For index_web_page, it defaults to "web_page_loader_v1" internally if rag_loader is "web_page" or not set.
        )
    )
    genie = await Genie.create(config=app_config)

    # Create a dummy directory and file for indexing
    data_path = Path("./rag_data_example")
    data_path.mkdir(exist_ok=True)
    (data_path / "sample.txt").write_text("Genie makes RAG easy and configurable.")

    # Index the directory
    collection = "my_sample_collection"
    await genie.rag.index_directory(str(data_path), collection_name=collection)
    print(f"Indexed documents from '{data_path}' into '{collection}'.")

    # Search the collection
    query = "How is RAG with Genie?"
    results = await genie.rag.search(query, collection_name=collection, top_k=1)
    if results:
        print(f"Search for '{query}':")
        for res in results:
            print(f"  - Score: {res.score:.4f}, Content: {res.content[:100]}...")
    else:
        print(f"No results found for '{query}'.")

    # Clean up dummy data
    (data_path / "sample.txt").unlink(missing_ok=True) 
    if data_path.exists(): data_path.rmdir() # Only rmdir if it's empty and exists

    await genie.close()

if __name__ == "__main__":
    asyncio.run(main())

Overriding RAG Component Configurations¶

You can override configurations for specific RAG component plugins (Document Loaders, Text Splitters, Embedding Generators, Vector Stores, Retrievers) using their respective *_configurations dictionaries in MiddlewareConfig.

For example, to configure the WebPageLoader to use Trafilatura for better content extraction when genie.rag.index_web_page() is called (and rag_loader feature is set to "web_page" or not set, allowing the default):

app_config = MiddlewareConfig(
    features=FeatureSettings(
        rag_loader="web_page", # Explicitly or implicitly uses web_page_loader_v1
        rag_embedder="sentence_transformer",
        rag_vector_store="faiss"
    ),
    document_loader_configurations={
        "web_page_loader_v1": { # Canonical ID of the WebPageLoader
            "use_trafilatura": True,
            # "trafilatura_include_comments": False # Other loader-specific settings
        }
    },
    embedding_generator_configurations={
        "sentence_transformer_embedder_v1": {
            "model_name": "paraphrase-multilingual-MiniLM-L12-v2" # Use a different ST model
        }
    }
)

When calling genie.rag.index_directory() or genie.rag.search(), you can also pass loader_id, splitter_id, embedder_id, vector_store_id, retriever_id, and their corresponding *_config dictionaries to override the defaults for that specific call.

Advanced Usage¶

The genie.rag interface internally uses the RAGManager. For advanced scenarios requiring direct interaction with the RAGManager or individual RAG component plugins, you can access it via genie._rag_manager (though this is generally not needed for typical use cases).

Refer to Creating RAG Plugins for information on developing your own RAG components.

Using Retrieval Augmented Generation (RAG)¶

Core RAG Operations via genie.rag¶

Configuring RAG Components¶

Overriding RAG Component Configurations¶

Advanced Usage¶

Core RAG Operations via `genie.rag`¶