Principle:Intel Ipex llm Embedding And Vector Store
| Knowledge Sources | |
|---|---|
| Domains | NLP, RAG, Embeddings |
| Last Updated | 2026-02-09 00:00 GMT |
Overview
Technique for generating document embeddings using a BGE model on Intel XPU and storing them in a Chroma vector database for similarity retrieval.
Description
Embedding and Vector Store creation involves two steps: (1) generating dense vector representations of text chunks using a BGE (BAAI General Embedding) model accelerated by IPEX-LLM on Intel XPU, and (2) inserting those embeddings into a Chroma vector database for fast similarity search. The IpexLLMBgeEmbeddings class provides XPU-optimized embedding generation, and Chroma provides the in-memory vector store with cosine similarity retrieval.
Usage
Use after document chunking to create a searchable vector store. The vector store serves as the retrieval component in the RAG chain, finding the most relevant chunks for a given query.
Theoretical Basis
# Abstract embedding + retrieval logic (NOT real implementation)
# 1. For each chunk c_i: embedding_i = BGE_model.encode(c_i) -> R^768
# 2. Store all (embedding_i, c_i) pairs in vector DB
# 3. For query q: q_emb = BGE_model.encode(q)
# 4. Retrieve top-K: argmin_i(cosine_distance(q_emb, embedding_i))