Fix: stop only at [RESULT:, inject results into assistant turn, ACTION: regex fix

Major fix: few-shot examples in system prompt, ACTION: format, stop tokens, 12-step loop

Fix: retry logic in web_search, tool loop prevention, cleaner system prompt

Upgrade duckduckgo-search to 7.5.3

Fix: sync LLM + async streaming, stop token for tool parsing, shorter prompts

Complete rewrite: FastAPI+SSE streaming, proper tool calling, web_search, token streaming

Switch to FastAPI+SSE, add duckduckgo-search, remove gradio

Use gradio 3.50.2 compatible app.py

Downgrade to gradio==3.50.2 + pydantic==1.x to fix jinja2 crash

fix: remove ssr_mode (not supported in gradio 4.44.1)

Fix: gr.Blocks with simple Chatbot, no type param, no ChatInterface

Pin gradio==4.42.0 (stable, no ssr_mode issue)

fix: requirements.txt (resolve RUNTIME_ERROR / gradio launch)

fix: app.py (resolve RUNTIME_ERROR / gradio launch)

fix: Dockerfile (resolve RUNTIME_ERROR / gradio launch)

Fix: use gr.ChatInterface instead of gr.Blocks+Chatbot to avoid Gradio 4.44 type bug

Fix: add type='tuples' to Chatbot, fix history iteration, remove share=True error

Update Dockerfile

Create requirements.txt

Update README.md

Update app.py

Update Dockerfile

Create app.py

Create Dockerfile

initial commit