fix: improve graph generation reliability (strip thinking blocks, add retry, robust DOT extraction)

fix: disable thinking mode for graph generation to fix corrupted DOT output

fix: switch to langchain-chroma + pin chromadb==0.5.23 to fix KeyError ephemeral

fix: use chromadb.EphemeralClient() to resolve KeyError ephemeral

refactor: slim down GRAPH_SYSTEM prompt (~65% token reduction)

Revert LLM back to nvidia/nemotron-3-nano-30b-a3b, update UI description

Switch LLM to google/gemma-4-31b-it, update all model descriptions in UI

Fix TTS synthesize() parameter name: use voice_name to match gRPC protocol

Fix TTS: use correct TTSService class, add fallback when TTS returns None

Reduce podcast narration length requirement to 200 words

Simplify podcast to single-voice narration: remove dual-speaker format from prompt and comments

Switch TTS from Google Gemini to NVIDIA Riva TTS (magpie-tts-multilingual hosted API)

Switch embedding model from Google to NVIDIA NIM (llama-nemotron-embed-1b-v2)

final

update name

update home

update prompt

Update RAG prompt: bullet points and sequential citations

Update dependencies: add NVIDIA endpoints

Sync HF metadata

without comment

update

add ignore

1.0

HF-Deployment: Initial commit