Swapping OpenAI Embeddings for a Local Qwen3 Model
Same pgvector table, same 1536 dimensions, nothing leaves the machine
My knowledge base embedded every note with OpenAI's text-embedding-3-small. I added Qwen3-Embedding-8B served by LM Studio as a second, fully local option, without a new table or a re-index. Here is what that took: Matryoshka truncation, an instruction prefix for queries, and a model_id column.
Posted by Jamie Craane on Monday, September 14, 2026
