ksoule/granite-switch-demo

Use requirements only in validation, never in the draft prompt

Restructure README around the two activities

Render adapter JSON verdicts upright, not italic

Append the IVR outcome beneath the checker bubble's KV note

Re-express judge prompts on the draft's decode tokens for KV reuse

Bold IVR outcome, code-styled KV notes in meta bubbles, all adapters default

Skip re-judging drafts identical to an already-failed attempt

Per-interaction KV prefix reuse with a ⚡ hit-rate note per generation

Merge IVR outcome into the checker bubble; reword exhaustion note

Checker bubble: verdict JSON first, then satisfied/not-satisfied note

Remove the per-message copy button

Meta-message styling: italic attempt labels, bare verdicts, purple notes

Label IVR draft bubbles with their attempt number

Render each assistant message as its own bubble

Default bubble styling; each adapter verdict in its own bubble

Stream draft tokens into the chat via mellea's async streaming

Validate IVR requirements on the live context (blind-judge workaround)

Show the adapters' JSON verdicts in the chat bubbles

Skip PEFT deactivation when no PEFT adapter was loaded

Flatten Gradio 6 block-format message content before instruct

Set python_version 3.12 (ZeroGPU max; mellea needs >=3.11)

Rewrite demo on Mellea's HF backend: embedded-adapter glue + IVR loop

Append user requirements to the original prompt (Requirements: template)

Trim roster to requirement-check, uncertainty, guardian-core

Revert to committed state (87cc118): Core+Guardian roster

Swap adapter roster: drop context-attribution/factuality-*, add RAG answerability/hallucination_detection/citations; package-prefixed names

Implement context-attribution's sentence-tagged protocol from the adapter README

Lighter purple for adapter responses; yellow chip for KV-cache notes

Carry each aLoRA adapter's invocation marker in its follow-up prompt

Multi-select adapters: each judges the first answer sequentially, sharing the prefix KV

Tint the whole adapter prompt/response bubbles purple; no inner background patches

Style adapter prompt purple/italic, move cache notes under responses, single-shot input

Reuse turn-1 decode tokens: extend the real sequence tensor instead of re-tokenizing

Publish KV-cache hit rate under each user message; reuse prefix KV for the adapter turn

Handle None from hidden Gradio textboxes

Two-turn flow: standard generation, then automatic Core/Guardian adapter invocation

Granite Switch 4.1 8B ZeroGPU demo

initial commit