Skip navigation

Engineering notes

The OpenVie blog

Posts that go into the hard parts of a self-hosted RAG system: fusing retrieval results, abstaining when evidence is missing, splitting OLTP from OLAP, approval workflow, and layered security. Every post links the sources it read so you can check them yourself.

Latest post

Latest post
Local modelsVietnameseOllama

A local Vietnamese stack: BGE-M3 and VyLinh 3B in OpenVie

A self-hosted RAG stack needs two models doing two different jobs: one embedder for retrieval, one generative model for writing the answer. OpenVie defaults to BGE-M3 and VyLinh 3B through Ollama, and this post explains both the choice and its limits.

Sections in this post
  1. Two different problems, two different models — why retrieval and generation have to split
  2. BGE-M3: three retrieval modes in one encoding pass
  3. Why 1024 dimensions with an 8192-token context matters for administrative text
  4. VyLinh 3B: a small model for Vietnamese
  5. The limits the model card itself admits
  6. Loading the models into Ollama

9 min read13 sections

Read post

All posts

Sources are listed at the end of each post