GPU Local Models

Supported path

Full local high-quality semantic retrieval is Linux-first and currently assumes:

• NVIDIA GPU
• conda available
• dedicated env: research_embed

Install

scibudy install-local-models

Warm model caches

scibudy warm-local-models --background

Current local model profile

• embedding: Qwen/Qwen3-Embedding-4B
• reranker: Qwen/Qwen3-Reranker-4B

Validation

scibudy analysis-settings
scibudy ingest-item <item_id> --reingest --skip-forums
scibudy search-evidence <library_id> calibration --format json

Expected signal:

• compute_backend: local_transformer
• semantic_backend: local_transformer+reranker