Text embeddings

RAG Pipeline & Semantic Search

• RAG Pipeline Goal: After text chunks are extracted, the system must find relevant chunks from the source document based on a user's query to use as context in the prompt.
• Core Problem: This process of finding related content is fundamentally a search problem.
• Solution: The most common implementation is Semantic Search.
• Mechanism: Semantic search utilizes Text Embeddings to understand the meaning of text and find the most related chunks to the user's question.

Text Embeddings

• Definition: A text embedding is a numerical representation of the meaning contained within a piece of text.
• Generation: Embeddings are generated by an Embedding Model.
• Output Structure: The model outputs a long list of numbers (the embedding).
• Value Range: The numbers within an embedding can range from negative 1 up to positive 1.
• Interpretation:
• Each number represents a score of some quality of the input text (e. g. , topic, sentiment).
• Crucial Note: The exact meaning of any specific number is not known; interpretations are helpful conceptual tools.

Implementation Details

• Tool: To generate embeddings, a dedicated Embedding Model (e. g. , Text Embedding 005) is used.
• Access: Access is managed via the Vertex SDK.
• Process: A helper function takes text input and passes it through the model to return the resulting embedding list.

Takeaways

• The RAG pipeline uses Semantic Search to find relevant text chunks from a source document based on a user's query.
• Semantic Search functions by utilizing Text Embeddings, which are numerical representations of the meaning within a piece of text.
• An Embedding Model generates these embeddings, outputting a list of numbers (e. g. , ranging from negative 1 to positive 1).
• Each number within an embedding represents a score for a specific quality of the input text, such as topic or sentiment.