RAG Pipeline

 

Documents

   ↓

Embeddings

   ↓

Vector DB (e.g., FAISS)

   ↓

Query → embedding

   ↓

Similarity search

   ↓

Top-k chunks

   ↓

LLM

   ↓

Answer

Retrieval-Augmented Generation (RAG) pipeline is divided into two main phases: the Ingestion phase (where data is prepared) and the Retrieval/Generation phase (where the user interacts with the data). 

RAG Workflow Architecture

• Ingestion Phase (Offline)
  • Documents: Raw data sources like PDFs or text files are loaded.
  • Embeddings: The text is broken into smaller pieces (chunks) and converted into numerical vectors using an embedding model.
  • Vector DB (e.g., FAISS): These vectors are stored in a specialized database that allows for fast mathematical searching.
• Inference Phase (Online)
  • Query → Embedding: When you ask a question, it is converted into a vector using the same embedding model used for the documents.
  • Similarity Search: The system compares the query vector against the vectors in the database to find the closest matches.
  • Top-k Chunks: The most relevant "k" pieces of text are retrieved from the database.
  • LLM → Answer: These chunks are sent to a Large Language Model (like GPT-4) as context, allowing it to generate an answer based on your specific documents. 

Visual Workflow Diagram

Below is a representation of how these components connect: