The three activated LoRA adapters qi loads at startup, in GGUF.
Hugging Face publishes these as safetensors under granitelib-rag-r1.0; these
are the GGUF conversions, which is why they live here rather than being mirrored
from upstream. The other two intrinsics — citations and hallucination_detection —
are plain LoRAs and are fetched from Hugging Face directly.
Activated means the adapter engages only after its invocation tokens, so the base
model's KV cache for the prompt stays valid. For answerability, which runs once
per source read, that is the difference between 16.98s and 0.04s per call.
Verified by sha256 on download against src/model/catalog.json:
rag-alora-answerability.gguf 819d5d10e3fbe2e7ebdce078585afb4477131d991d1ed4a11ce0bbc421273502
rag-alora-query_clarification.gguf 9bb08cd41de220424e6cfcd5a2095643831a924cb5b4b7c742784b033360c796
rag-alora-query_rewrite.gguf c312f60f44f6b3997824cadf912b7aa39d7770eb9cd1c7c7589b3a5bc26201f7
Base model: ibm-granite/granite-4.1-3b. Apache-2.0, as upstream.