✅ Real-time Search: DuckDuckGo search engine, stable and reliable
✅ AI Summarization: Automatically reads web pages and extracts key information
✅ Multi-LLM Support: GPT-5.1, DeepSeek-V3, Gemini 3 Pro, Claude 4.5, Qwen3, and 8+ LLM providers
✅ Modular Architecture: Separated API config, search settings, and execution logic
✅ Chinese-Friendly: Perfect support for Chinese LLM providers (Qwen, Doubao, DeepSeek)

Typical Use Cases:

🌤️ Query real-time weather to generate scene images
📰 Fetch latest news and generate related content
🔍 Fact checking and information verification
🛍️ Product information lookup and review summarization
📍 GPS coordinate to location conversion (e.g., "40.00023, 116.27808" → "Beijing, Haidian District")
💬 Pure LLM chat (disable web search for direct AI conversations)
🌐 Any creative workflow requiring internet information

🏗️ New Modular Architecture

Inspired by comfyui_LLM_party's excellent design, we adopt a modular layered architecture:

📊 Node Composition

🔑 API Loader → ⚙️ Settings → 🌐 Search Agent → Results

Node	Function	Output
🔑 Live Search API Loader	API config & model selection	MODEL_CONFIG
⚙️ Live Search Settings	Search parameters	SEARCH_SETTINGS
🌐 Live Search Agent	Main search logic	answer, source_urls, optimized_prompt

✅ New Architecture Benefits

Modular Design: Separation of config and logic, easier to maintain
Reusability: One API Loader can connect to multiple Agents
Flexibility: Different Settings for different scenarios
Professional: Best practices from large-scale projects

🔄 Backward Compatibility

Legacy single-node 🌐 Live Search (Legacy) still available
New users recommended to use the new three-node combo

✨ Key Features

🔍 DuckDuckGo Search Engine
Stable, proxy-friendly, no API key required. Perfect for real-time info retrieval and complements any LLM/VLM.
🧠 Multi-Provider Model Hub (Text & Vision)
- OpenAI: GPT-5.x, GPT-4.1 family, GPT-4o (now fully TI2T-capable)
- SiliconFlow: DeepSeek, Qwen3/Qwen2.5 VL, GLM-4.5V, Qwen3-Omni, etc. (single provider covering 70+ CN models)
- DeepSeek Official / Aliyun / Volcengine, Gemini (OpenAI-format), Anthropic Claude, Qwen (Aliyun), Grok, Doubao, Ollama (local)
✅ Every provider is validated end-to-end (auth, payload, response parsing).
🖼️ Dual Mode: T2T & TI2T
- T2T: Traditional text-only LLM flow with web search + summarization
- TI2T: Vision-language pipeline (prompt + IMAGE) for reverse image search, landmark recognition, document reading, etc.
- TI2T supports web search + image context fusion (two-stage VLM call) and is available for OpenAI GPT-5.1 / GPT-4o families & SiliconFlow VLMs.
🌦️ Structured Tooling
- Open-Meteo integration for precise real-time weather/time when GPS coordinates are provided
- geopy + Nominatim reverse geocoding (lat/lon → city/country)
- Query Optimization (rename: optimize_query) improves English search keywords while keeping answers in your selected language
- Conflict resolution prompts ensure VLM trusts web data for up-to-date info (time, weather).
🈚 Multi-Language Output
Explicit output control: 中文 or English (no more “Auto” surprises). System prompts enforce consistency for both LLM and VLM flows.
🔐 Cloud-Friendly Security
API keys never persist to disk, .env + api_config.json supported, and per-node proxy keeps shared GPU rentals safe.

🧠 T2T vs TI2T Modes

Mode	Inputs	Model Slot	Flow	Perfect For
T2T (Text → Text)	Prompt	`t2t_model`	Search (DuckDuckGo) → scrape trusted sources → inject Open-Meteo & geopy context → LLM summarization.	Weather/time, news, fact checking, research, pure LLM chat (disable web search).
TI2T (Text + IMAGE → Text)	Prompt + ComfyUI IMAGE tensor	`ti2t_model`	Encode image → VLM query generation → DuckDuckGo search → VLM final reasoning with image + search snippets.	Landmark recognition, “reverse image” info, screenshot comprehension, visual + real-time hybrid tasks.

Implementation Notes

API Loader now provides independent t2t_model / ti2t_model dropdowns per provider. Agent auto-selects based on Settings mode.
TI2T currently supports OpenAI GPT-5.1 / GPT-4o / GPT-4-turbo and SiliconFlow VLM lineups (Qwen3-VL, Qwen3-Omni, GLM-4.5V, DeepSeek-VL2, etc.).
TI2T keeps web search optional: when enabled, VLM runs twice (query generation + final answer) so image context stays synced with live data.
Both modes reuse the unified MODEL_CONFIG output, so legacy workflows remain compatible.

OpenAI GPT-5.x usage notes

Access to GPT-5.1 family still requires an OpenAI plan with GPT-5 privileges (Pro/Team/Enterprise). If the API returns 404/403, double-check your account permissions.
Live Search automatically switches GPT-5.* calls to the Responses API. You do not need to edit the base URL; simply pick gpt-5.1, gpt-5.1-mini, gpt-5, gpt-5-mini, or gpt-5-pro in API Loader.
The node maps max_tokens → OpenAI's max_output_tokens under the hood, so you can keep using the same control in the UI.
TI2T (image + text) is currently validated for gpt-5.1, gpt-5.1-mini, gpt-5, gpt-5-mini, and gpt-5-pro. Older GPT-4 models stay available for backward compatibility.

🚀 Installation

Method 1: via ComfyUI Manager (Recommended)

Open Manager panel in ComfyUI
Click Install Custom Nodes
Search for Live Search
Click Install and restart ComfyUI

Method 2: Git Clone

Navigate to your ComfyUI custom_nodes directory and run:

git clone https://github.com/Zone-Roam/ComfyUI-Live-Search.git
cd ComfyUI-Live-Search

# If using Portable version of ComfyUI (Recommended)
..\..\..\python_embeded\python.exe -m pip install -r requirements.txt

# If using system Python or virtual environment
pip install -r requirements.txt

Then restart ComfyUI.

Method 3: Manual Installation

Download the ZIP file
Extract it to ComfyUI/custom_nodes/ComfyUI-Live-Search
Install dependencies as shown in Method 2
Restart ComfyUI

🛠️ Usage Guide

📸 Workflow Example

The image below shows both usage methods:

Left: Legacy single-node mode - Simple and fast, all configs in one node
Bottom right: New modular architecture - API Loader + Settings + Agent, more flexible and reusable

Method 1: New Modular Architecture (Recommended) ⭐

1. 🔑 Live Search API Loader

Configure LLM API and model parameters.

Parameter	Description
provider	Choose provider: OpenAI, SiliconFlow, DeepSeek, Gemini, Anthropic, Qwen, Doubao, Ollama, etc.
t2t_model	Text-only model for T2T mode (LLM)
ti2t_model	Vision-language model for TI2T mode (VLM). Shows “No VLM models available” if the provider has none.
api_key	API key (optional, supports .env)
base_url	API endpoint (optional, falls back to provider defaults)
temperature	Temperature (0.0-2.0)
max_tokens	Maximum output length
timeout	Request timeout

2. ⚙️ Live Search Settings

Configure search behavior.

Parameter	Description
mode	`T2T` (text) or `TI2T` (text + image). TI2T expects an IMAGE input on the Agent node.
enable_web_search	Enable/disable web search (OFF = use as pure LLM)
num_results	Number of search results (1-10)
output_language	Output language: `中文` or `English`
optimize_query	LLM-powered search keyword optimization (English-focused for better search recall)
proxy	Proxy address (optional)

3. 🌐 Live Search Agent

Main search node, connects to the above two nodes.

Input	Type	Description
prompt	STRING	Your question
model_config	MODEL_CONFIG	From API Loader
search_settings	SEARCH_SETTINGS	From Settings
(optional) image	IMAGE	Required when `mode = TI2T`. Pass any ComfyUI image tensor (RGB/RGBA).
(optional) role	STRING	Custom system prompt injected before the default instructions.

Output	Description
answer	AI-generated answer
source_urls	Referenced source links
optimized_prompt	Optimized search query

Method 2: Legacy Single-Node Mode

Node: 🌐 Live Search (Legacy)

Input Parameters

Parameter	Description
prompt	Your question. Supports both Chinese and English. e.g., "What's the weather in Beijing?" or "北京现在的天气"
output_language	🌐 Output Language • 中文: Force Chinese output • English: Force English output
optimize_query	🔄 Search Keyword Optimization (Recommended ON) • OFF (default): Use original input directly • ON: LLM rewrites the query (English keywords) for better DuckDuckGo recall - Preserves answer language via Agent settings - Removes redundancy, keeps core info - Outputs before/after comparison
provider	Choose your LLM provider: `OpenAI`, `DeepSeek (Official/Aliyun/Volcengine)`, `Gemini`, etc.
model	🎯 Model Selection (Dropdown) • OpenAI: gpt-5.1, gpt-5, gpt-4.1, o3, o3-pro, etc. • DeepSeek: deepseek-v3, deepseek-chat, deepseek-reasoner • Gemini: gemini-3-pro, gemini-2.5-pro, gemini-2.5-flash, etc. • Claude: claude-sonnet-4-5, claude-haiku-4-5, etc. • Supports search filtering for quick model lookup
api_key	(Optional) Your API Key. If left empty, it tries to load from config files.
proxy	(Optional) Proxy address like `http://127.0.0.1:7890`. Leave empty for direct connection.

Outputs

Output	Description
answer	AI-generated answer based on search results
source_urls	List of referenced web page links
optimized_prompt	Shows prompt optimization status (whether optimized, before/after comparison)

Example Workflows

1. Real-time Weather Image Generation

Input: "What's the weather in Beijing?"
Optimize: ON ✅
Optimized: "Beijing weather now"
Output: "Currently 2:00 PM in Beijing, Sunny, 15°C."
→ [Connect to Text2Image] → Generate Beijing sunny street scene

2. Fact Checking

Input: "Who won the latest Super Bowl?"
Optimize: ON ✅
Output: Accurate answer based on real-time results

3. Cross-Language Query

Input: "北京现在的天气" (Chinese question)
Output Language: English 🇺🇸
Optimize: ON ✅
Output: Beijing weather info (answered in English)

4. International Collaboration

Input: "What's the weather in Beijing?" (English question)
Output Language: 中文 🇨🇳
Output: Beijing weather info (answered in Chinese)

5. GPS Coordinate Search

Input: "What's the weather at coordinates 40.00023, 116.27808?"
Optimize: ON ✅
Auto-conversion: Coordinates → "Beijing, Haidian District"
Optimized Query: "timeanddate Beijing Haidian"
Output: Real-time weather and time for Haidian District, Beijing

6. Pure LLM Mode (No Search)

Settings: enable_web_search = OFF
Input: "Explain quantum computing"
Output: Direct LLM response without web search (faster, no internet required)

🔍 Why Only DuckDuckGo?

This node uses real web scraping for search, not API calls. In our testing:

✅ DuckDuckGo Advantages:

Automation-friendly with lenient anti-bot measures
Works reliably even with proxy configuration
Search quality fully meets real-time information retrieval needs
Open-source friendly with strong community support

❌ Google Issues:

Extremely strict anti-scraping mechanisms (CAPTCHAs, IP blocks, User-Agent detection)
Often returns empty results or CAPTCHA pages even with proxies
googlesearch-python library is unstable in production
Frequent access leads to temporary IP bans

💡 If You Need Google Search Quality:

Consider using official Google Custom Search API (paid)
Or use third-party services like SerpAPI (paid)

We chose DuckDuckGo to ensure the node works reliably across all environments.

💰 Free Usage Options

Don't want to pay for APIs? Here are completely free solutions!

Recommended: Ollama Local Models ⭐

Install Ollama: Visit https://ollama.com/ to download

Download Models:

ollama pull llama4       # Recommended: Meta Llama 4 (2025 Latest)
ollama pull qwen3        # Or: Alibaba Qwen3
ollama pull deepseek-v3  # Or: DeepSeek V3

Configure in Node:
- Provider: Ollama (Local)
- Model: llama4 (or other downloaded models: qwen3, deepseek-v3, phi4)
- Base URL: http://localhost:11434/v1
- API Key: Leave empty

Advantages:

✅ Completely free, no usage limits
✅ Privacy-focused, runs entirely locally
✅ No network latency (except search part)
✅ Supports latest models like llama4, qwen3, deepseek-v3, phi4

Alternative: Ultra-Low-Cost APIs

If local resources are limited, consider these nearly-free options:

Provider	Price	Notes
DeepSeek	$0.14/million tokens	Official API, best value
SiliconFlow	Free tier available	China-friendly
Groq	Free tier available	Extremely fast

⚙️ Configuration (Optional)

For local users who don't want to paste their API key every time, there are two configuration methods:

Method 1: Use .env File (Recommended) ⭐

Copy .env.example to .env
Edit .env and fill in your API keys:

OPENAI_API_KEY=sk-your-openai-key-here
DEEPSEEK_OFFICIAL_API_KEY=sk-your-deepseek-key-here

Advantages:

✅ Industry standard practice
✅ Automatically excluded by .gitignore, won't be accidentally committed
✅ More secure and professional

Method 2: Use api_config.json

Rename api_config_example.json to api_config.json
Edit and fill in your API keys:

{
    "openai_api_key": "sk-...",
    "deepseek (official)_api_key": "sk-..."
}

API Key Priority

Node Input (Highest) > .env File > api_config.json (Lowest)

Note

: On cloud platforms, always use the api_key widget in the node for security.

📄 License

Apache 2.0 License