[Feature Request] AI chat/RAG: cache the embedding model instead of reloading it on every query

The problem load_or_build_index (src/paperless_ai/indexing.py) calls get_embedding_model(config) on every invocation, and that runs once per chat/RAG query (_stream_chat_with_documents in src/paperless_ai/chat.py, inside the read_store() block). There is no caching of the embedding instance, so each query reloads the full embedding model before it can retrieve anything. What we measured Per-query model reload cost on a production archive (about 1,050 documents, HuggingFace/CPU backend), by embedding model: | Embedding model | RAM loaded per query | |---|---| | intfloat/multilingual-e5-small | ~0.5 GB | | intfloat/multilingual-e5-base | ~1.1 GB | | BAAI/bge-m3 | ~2.3 GB | The reload dominates end-to-end query latency (it outweighs the actual vector search) and produces a RAM spike on every request. On memory-constrained instances this effectively caps the usable embedding model size for reasons that have nothing to do with retrieval quality. Suggested direction …