RAG+RAGAS+LangChain+FAISS+OpenAI

[Lomanu4]

RAG (Retrieval-Augmented Generation) Workflow

Import Required Libraries


This imports all the necessary libraries for loading datasets, performing text splitting, creating embeddings, building a retrieval system, and evaluating metrics.

dotenv: For loading environment variables (e.g., API keys).

load_diabetes: Provides the diabetes dataset from scikit-learn.

LangChain libraries: Tools for text splitting, embeddings, and setting up a question-answering (QA) pipeline.

FAISS: A library for efficient similarity search and clustering of dense vectors.

Ragas: For evaluating retrieval-based question-answering systems.

Dataset: From the datasets library, useful for organizing data for evaluation.

userdata: Used for securely retrieving sensitive data in Google Colab.

1. Ground Truth - Source of Truth


The foundation of the system lies in the source data. In this example:


diabetes = load_diabetes()
raw_text = diabetes.DESCR

• load_diabetes(): Fetches the diabetes dataset from sklearn.datasets.
• diabetes.DESCR: Contains a detailed description of the dataset (variables, data characteristics).
• raw_text: Represents the "Ground Truth" that the RAG system will reference for its operations.

2. Retrieval - Finding Relevant Information


The raw text (input) is split into smaller chunks, and a vector store is created using FAISS. Queries retrieve the most relevant chunks.


# Split into chunks (simulate document retrieval)
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50)
texts = text_splitter.split_text(raw_text)

# Create Embeddings & Build FAISS Index
embeddings = OpenAIEmbeddings(model="text-embedding-ada-002") # Recommended model
docsearch = FAISS.from_texts(texts, embeddings)

• RecursiveCharacterTextSplitter: Splits the large raw_text into smaller, overlapping segments (texts).
• OpenAIEmbeddings: Converts input text chunk into numerical vector representations for processing.
• FAISS.from_texts(texts, embeddings): Builds an index of these embeddings for efficient retrieval.

3. Augmentation - Adding Context


The RetrievalQA chain in LangChain manages augmentation.

• It takes the user's query and uses the docsearch index to find relevant documents.
• These documents are passed to the LLM along with the original query, providing contextual information.

llm = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0)
qa_chain = RetrievalQA.from_chain_type(
llm=llm,
chain_type="stuff",
retriever=docsearch.as_retriever(),
return_source_documents=True # helpful for debugging
)

• RetrievalQA.from_chain_type sets up the RAG pipeline.
• retriever=docsearch.as_retriever() connects FAISS index to QA chain, enables fetch relevant doc/info based on query
• chain_type="stuff" pass retrieved doc/info to LLM. It stuffs query+info/doc.
• return_source_documents=True for traceability

4. Generation - Producing the Response


The LLM generates a response by combining the query and augmented context retrieved in the previous step.


for query in queries:
result = qa_chain.invoke({"query": query})
answers.append(result["result"])

# Extract retrieved docs for Ragas evaluation
retrieved_docs = result.get("source_documents", [])
contexts.append([doc.page_content for doc in retrieved_docs])
5. Traceability - Explaining the Source


The system ensures traceability by showing the origin of the retrieved information, helping users understand the response's basis.


retrieved_docs = result.get("source_documents", [])
contexts.append([doc.page_content for doc in retrieved_docs])

Пожалуйста Авторизируйтесь или Зарегистрируйтесь для просмотра скрытого текста.