LocalRAG v0.10.0

LocalRAG 0.10.0 refreshes the local model stack while keeping the proven default configuration conservative.

Highlights

Recommended Ollama models are grouped into practical hardware profiles:
- qwen3.5:4b — light
- qwen3.5:9b — balanced default
- gemma4:12b — quality
- granite4.2:8b — experimental RAG profile
- qwen3.6:27b — experimental workstation profile
Recommended embedding models now include Qwen/Qwen3-Embedding-0.6B and optional google/embeddinggemma-300m.
E5 uses the required query: and passage: retrieval prefixes.
Qwen3 Embedding uses its query instruction; EmbeddingGemma uses retrieval query/document prompts; GTE enables its required remote model code.
All supported embedding profiles normalize vectors.
FAISS persists the embedding model and adapter version and refuses incompatible legacy indexes, triggering a safe rebuild.
The model manager shows download size, context window, hardware profile, stability status, and license caveats.
The bundled Ollama image is updated to 0.33.3.

The default runtime remains qwen3.5:9b with intfloat/multilingual-e5-large.
The first start after upgrading rebuilds a legacy FAISS index because E5 now uses its correct asymmetric retrieval prefixes.
google/embeddinggemma-300m requires accepting the Gemma license while signed in to Hugging Face.
granite4.2:8b is marked experimental for Russian and Hebrew because they are not listed among its officially supported languages.
Older installed Ollama models remain selectable and participate in fallback resolution; they are only removed from the primary recommendation list.