text-embedding-3-small vs text-embedding-ada-002 vs BGE-large-en-v1.5: Which Embedding Model for RAG?
Practical comparison of OpenAI text-embedding-3-small, text-embedding-ada-002, text-embedding-3-large, BGE-large-en-v1.5, and BGE-M3 for RAG pipelines in LangChain, LlamaIndex, Pinecone, and Weaviate — covering MTEB scores, token limits, and cost.
Choosing an embedding model for a Pinecone, Weaviate, or pgvector RAG pipeline involves three variables: retrieval quality (MTEB benchmark), token limit, and cost per million tokens. text-embedding-ada-002 (OpenAI) — the most widely deployed model in production LangChain and LlamaIndex pipelines — produces 1,536-dimensional vectors, has an 8,191-token limit, and costs $0.10/1M tokens. text-embedding-3-small supersedes it on every metric: higher MTEB BEIR scores, same 1,536 dimensions, same token limit, and costs $0.02/1M tokens — a 5× cost reduction with better retrieval quality. text-embedding-3-large adds a 3,072-dimension option at $0.13/1M tokens for use cases where ranking precision matters more than cost. For teams that want free, self-hosted embeddings, BGE-large-en-v1.5 (BAAI, 512-token limit) consistently outperforms ada-002 on BEIR benchmarks and runs in-process via sentence-transformers or Transformers.js without API calls. BGE-M3 extends this to multilingual retrieval with a 8,192-token limit, making it the default choice for non-English corpora.
The model you pick interacts directly with your chunking strategy: a 512-token BGE-large limit means chunks larger than ~380 words get silently truncated before embedding, ruining retrieval for long passages. The RAG Chunk Inspector lets you paste a document, set chunk size and overlap, and preview each chunk's token count against the model's limit — so you can confirm no chunks exceed the embedding boundary before indexing into Pinecone or Weaviate.
Follow Trango Compute on LinkedIn
We post updates on new tools, context engineering patterns, and LLM cost research.