Skip to content

RAG and RAGConfig classes

The RAG class

The RAG class orchestrates the components necessary for a retrieval-augmented generation pipeline. It initializes with a configuration, either directly or from a file.

RAG

The RAG object has two main purposes:

  • loading the RAG with documents, which involves ingesting and processing documents to be retrievable by the system
  • generating the chain from the components as specified in the configuration, which entails assembling the various components (language model, embeddings, vector store) into a coherent pipeline for generating responses based on retrieved information.

Loading and querying documents

from pathlib import Path
from backend.rag_components.rag import RAG

rag = RAG(config=Path(__file__).parent / "backend" / "config.yaml")
chain = rag.get_chain()

print(chain.invoke("Who is bill Gates?"))
# > content='Documents have not been provided, and thus I am unable to give a response based on them. Would you like me to answer based on general knowledge instead?'

rag.load_file(Path(__file__).parent / "data_sample" / "billionaires.csv")
# > loader selected CSVLoader for /.../data_sample/billionaires.csv
# > {'event': 'load_documents', 'num_added': 2640, 'num_updated': 0, 'num_skipped': 0, 'num_deleted': 0}

print(chain.invoke("Who is bill Gates?"))
# > content='Bill Gates is a 67-year-old businessman from the United States, residing in Medina, Washington. He is the co-chair of the Bill & Melinda Gates Foundation and is recognized for his self-made success, primarily through Microsoft in the technology industry. As of the provided document dated April 4, 2023, Bill Gates has a final worth of $104 billion, ranking him 6th in the category of Technology. His full name is William Gates, and he was born on October 28, 1955.'

RAGConfig

Configuration of the RAG is done using the RAGConfig dataclass. You can instanciate one directly in python, but the preferred way is to use the backend/config.yaml file. This YAML is then automatically parsed into a RAGConfig that can be fed to the RAG class.

The configuration provides you with a way to input which implementation you want to use for each RAG components:

  • The LLM
  • The embedding model
  • The vector store / retreiver
  • The memory / database

Zooming in on the LLMConfig as an example: 

@dataclass
class LLMConfig:
    source: BaseChatModel | LLM | str
    source_config: dict
    temperature: float

  • source is the name of name of the langchain class name of your model, either a BaseChatModel or LLM.
  • source_config is are the parameters used to instanciate the source.
  • temperature regulates the unpredictability of a language model's output.

Example of a configuration that uses a local model served with Ollama. In backend/config.yaml: 

LLMConfig: &LLMConfig
  source: ChatOllama
  source_config:
    model: tinyllama
    temperature: 0

Configuration recipes

You can find fully tested recipes for LLMConfig, VectorStoreConfig, EmbeddingModelConfig, and DatabaseConfig in the Cookbook.

This is the python equivalent that is generated and executed under the hood when a RAG object is created. 

llm = ChatOllama(model="tinyllama", temperature=0)

You can also write the configurations directly in python, although that's not the recommended approach here. 

from langchain_community.chat_models import ChatOllama

from backend.config import LLMConfig

llm_config = LLMConfig(
    source=ChatOllama,
    source_config={"model": "llama2", "temperature": 0},
)

Extending the RAGConfig

See: How to extend the RAGConfig