user can define search timeout

give 5 second for web earch to gather reults

support thinking models and streamingly display thought

do not preview prompt at error return from chat response

inject assistant placeholder at right time

disable L142 which is not needed

fix bug in prompt preview display

add prompt preview for debug

fix: prevent self-talking issue by using tokenizer chat_template formatting

bugfix to Error: "str" object has no attribute "pad_token_id"

add taiwan elm 1.1b & 270m instruct

add type in qwen3 0.6b repo id

add qwen3

Add Smollm2 360m instruct fine-tuned on TaiwanChat

keep debug message

add debug to show web resarch result

give 1 second for web search to grab data

inject web search result if web search enabled

refactor(app): improve streaming, background search, dtype fallback, and cleanup :contentReference[oaicite:0]{index=0}

bugfixc: not using pipeline for response generation

Add original SmolLM2 135M Instruct for comparaison

Add SmolLM2-135M-Instruct-TaiwanChat

Add SmolLM2-135M TaiwanChat

Update README.md

default to gemma-3-4b

model repo_id typo fix

enable web search by default

remove tinyllama which has bad response quality

make streaming response

apply history flatten before it goint to prompt

better management on system prompt

add accelerate

usue chat pipeline instead of model and tokenizer individually

bugfix to padding-related issues

add attention mask

Clean model description

pin torch to 2.4.0

add sentencepiece tokenzier

update requirements

enable zerogpu with decorator

Apply ZeroGPU

add apt.txt

adjust requirements

adjust requirements.txt

specify gpu build for llama cpp python

Update README.md

fix readme

switch to gradio version for stability reason

add taiwan tinyllama

Improve responsiveness by asynchronously retrieving web search context