apps/gradio-space/README.md

gradio-space

Build Small hackathon UI — custom Studio frontend plus Classic Gradio tabs.

Entry points

URL	What
/	Studio UI — custom HTML/CSS/JS served via gradio.Server (near parity with Classic)
/classic	Full Gradio Blocks app (fallback / power-user tabs)

uv run --package gradio-space python -m gradio_space.server
# or
uv run --package gradio-space python -m gradio_space.app

Off Brand architecture

This package uses Gradio 6 Server mode (gradio.Server):

• Custom routes: GET /, static assets at /static/studio/
• API endpoints via @server.api(name=...) — callable from @gradio/client and gradio_client
• Classic UI mounted with mount_gradio_app(..., path="/classic")

Studio API names

Research & slides

• list_sessions, list_documents, session_memory
• discover_sources, auto_search_ingest, ingest_sources, ingest_url, ingest_files
• research_chat, generate_slides (supports source_mode: none / web / rag)
• generate_slides_from_conversation — build a deck from Research, Language lessons, or Chat history
• generate_quiz — printable MCQ worksheet (DOCX + HTML) with optional RAG / web sources

Voice & coach

• language_lesson_turn — unified text/audio turn for Language lessons (mode, language, auto_voiceout, coach variant)
• teacher_voice_turn, teacher_voice_audio_turn, teacher_voice_clear, teacher_voice_speak
• load_sample_pitch, analyze_pitch (Classic EchoCoach; language, ASR preset, speak_rewrite)
• recording_status, recording_start, recording_stop
• voice_presets

Settings & debug

• model_status, model_choices, set_active_model, reload_model
• debug_chat
• save_upload

Switching models locally

Set ALLOW_MODEL_SWITCH=true in .env (see USAGE.md). The Settings drawer and Classic Settings accordion share one runtime preset — changing it reloads weights and applies to Lesson slides, Research, and voice tabs (not just Chat debug).

Preset	Backend
minicpm-v-4.6	transformers (full VLM)
minicpm-v-4.6-gguf	llama.cpp (Llama Champion track)
minicpm5-1b	transformers

Demo script (judges) — teaching loop

1. Open / — Small Model Finetuning project workspace
2. Research — ingest a PDF or URL on your topic → ask 2 RAG questions with citations
3. Tap Generate slides from chat → switch to Slides → preview deck → Present (fullscreen, arrow keys)
4. Tap Create quiz on this topic → Quiz view → generate worksheet → download DOCX (answer key included)
5. Download PPTX and expand Agent trace
6. Optional: Language lessons → French voice turn → Slides from chat on the same topic

Classic UI (/classic) adds a Quiz maker tab after Lesson slides with the same agent pipeline.

Language lessons + Cohere stack (voice demo)

Badge line: Cohere Labs — Transcribe + Tiny Aya on one local Language lessons page.

1. Open / — Small Model Finetuning project workspace
2. Language lessons tab → select French → hold mic → ask « Explique le fine-tuning en termes simples. » → hear Piper/VibeVoice reply
3. Switch to Spanish, type a follow-up (text in, text + audio out with Auto-speak replies on)
4. Select Other (text only) → enter hi → show Tiny Aya Fire-quality written lesson (text only banner)
5. Toggle Use indexed sources after ingesting one PDF in Research
6. Optional: Classic UI (/classic) for EchoCoach pitch metrics

Space secrets for GPU demo:

ECHOCOACH_ASR_PRESET=cohere-transcribe
ECHOCOACH_COACH_MODEL=tiny-aya-global
ECHOCOACH_REALTIME_TTS_PRESET=vibevoice-realtime-0.5b

Space card metadata lives in the repository root README.md.