Hey folks! 👋 Been diving deep into building some RAG pipelines lately and I'm trying to shift away from the OpenAI embedding API for cost and control reasons. I keep hearing great things about the BGE (BAAI General Embedding) models, especially `BAAI/bge-small-en-v1.5` and the larger variants.
I'm curious if anyone in the community is running LlamaIndex with these open-source embeddings in production, or even in a serious staging environment.
* What's your setup? Are you using the `HuggingFaceEmbedding` integration, or something more custom?
* How's the performance and accuracy compared to `text-embedding-ada-002` in your tests? I've run some basic benchmarks on MTEB, but real-world retrieval feel is different.
* Any major pitfalls or "gotchas" with dimensions, normalization, or the query instruction prefixes that BGE models recommend?
* What's your chunking strategy paired with these models? I found I had to tweak my chunk size a bit from my OpenAI setup.
I'm currently testing with the `bge-base-en` model and a simple Pinecone vector store. Initial results are promising on my data, but I'd love to hear about your experiences before doubling down on this stack. The idea of having a solid, self-hosted embedding model is really appealing for scaling a lot of automated campaign content analysis I want to do.
Cheers, looking forward to the discussion!
Always A/B test.