README.md

metadata

title: CognitiveKernel-Launchpad
emoji: 🧠
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.44.1
app_file: app.py
pinned: false
license: mit
hf_oauth: true
hf_oauth_expiration_minutes: 480

🧠 CognitiveKernel-Launchpad — Hugging Face Space

This Space hosts a Gradio UI for CognitiveKernel-Launchpad and is tailored for Hugging Face Spaces.

• Original project (full source & docs): https://github.com/charSLee013/CognitiveKernel-Launchpad
• Access: Sign in with Hugging Face is required (OAuth enabled via metadata above).

🔐 Access Control

Only authenticated users can use this Space. Optionally restrict to org members by adding to the metadata:

hf_oauth_authorized_org: YOUR_ORG_NAME

🚀 How to Use (in this Space)

1. Click “Sign in with Hugging Face”.
2. Ensure API secrets are set in Space → Settings → Secrets.
3. Ask a question in the input box and submit.

🔧 Required Secrets (Space Settings → Secrets)

• OPENAI_API_KEY: your provider key
• OPENAI_API_BASE: e.g., https://api-inference.modelscope.cn/v1/chat/completions
• OPENAI_API_MODEL: e.g., Qwen/Qwen3-235B-A22B-Instruct-2507

Optional:

• SEARCH_BACKEND: duckduckgo | google (default: duckduckgo)
• WEB_AGENT_MODEL / WEB_MULTIMODAL_MODEL: override web models

🖥️ Runtime Notes

• CPU is fine; GPU optional.
• Playwright browsers are prepared automatically at startup.
• To persist files/logs, enable Persistent Storage (uses /data).

—

🧠 CognitiveKernel-Launchpad — Open Framework for Deep Research Agents & Agent Foundation Models

🎓 Academic Research & Educational Use Only — No Commercial Use 📄 Paper (arXiv:2508.00414) | 🇨🇳 中文文档 | 📜 LICENSE

---

🌟 Why CognitiveKernel-Launchpad?

This research-only fork is derived from Tencent's original CognitiveKernel-Pro and is purpose-built for inference-time usage. It removes complex training/SFT and heavy testing pipelines, focusing on a clean reasoning runtime that is easy to deploy for distributed inference. In addition, it includes a lightweight Gradio web UI for convenient usage.

---

🚀 Quick Start

1. Install (No GPU Required)

git clone https://github.com/charSLee013/CognitiveKernel-Launchpad.git
cd CognitiveKernel-Launchpad
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Set Environment (Minimal Setup)

export OPENAI_API_KEY="sk-..."
export OPENAI_API_BASE="https://api.openai.com/v1"
export OPENAI_API_MODEL="gpt-4o-mini"

3. Run a Single Question

python -m ck_pro "What is the capital of France?"

✅ That’s it! You’re running a deep research agent.

---

🛠️ Core Features

🖥️ CLI Interface

python -m ck_pro \
  --config config.toml \
  --input questions.txt \
  --output answers.txt \
  --interactive \
  --verbose

Flag	Description
-c, --config	TOML config path (optional)
-i, --input	Batch input file (one Q per line)
-o, --output	Output answers to file
--interactive	Start interactive Q&A session
-v, --verbose	Show reasoning steps & timing

---

⚙️ Configuration (config.toml)

TOML > Env Vars > Defaults

Use the examples in this repo:

• Minimal config: config.minimal.toml — details in CONFIG_EXAMPLES.md
• Comprehensive config: config.comprehensive.toml — full explanation in CONFIG_EXAMPLES.md

🚀 Recommended Configuration

Based on the current setup, here's the recommended configuration for optimal performance:

# Core Agent Configuration
[ck.model]
call_target = "https://api-inference.modelscope.cn/v1/chat/completions"
api_key = "your-modelscope-api-key-here"  # Replace with your actual key
model = "Qwen/Qwen3-235B-A22B-Instruct-2507"

[ck.model.extract_body]
temperature = 0.6
max_tokens = 8192

# Web Agent Configuration (for web browsing tasks)
[web]
max_steps = 20
use_multimodal = "auto"  # Automatically use multimodal when needed

[web.model]
call_target = "https://api-inference.modelscope.cn/v1/chat/completions"
api_key = "your-modelscope-api-key-here"  # Replace with your actual key
model = "moonshotai/Kimi-K2-Instruct"
request_timeout = 600
max_retry_times = 5
max_token_num = 8192

[web.model.extract_body]
temperature = 0.0
top_p = 0.95
max_tokens = 8192

# Multimodal Web Agent (for visual tasks)
[web.model_multimodal]
call_target = "https://api-inference.modelscope.cn/v1/chat/completions"
api_key = "your-modelscope-api-key-here"  # Replace with your actual key
model = "Qwen/Qwen2.5-VL-72B-Instruct"
request_timeout = 600
max_retry_times = 5
max_token_num = 8192

[web.model_multimodal.extract_body]
temperature = 0.0
top_p = 0.95
max_tokens = 8192

# Search Configuration
[search]
backend = "duckduckgo"  # Recommended: reliable and no API key required

🔑 API Key Setup

1. Get ModelScope API Key: Visit ModelScope to obtain your API key
2. Replace placeholders: Update all your-modelscope-api-key-here with your actual API key
3. Alternative: Use environment variables:

   export OPENAI_API_KEY="your-actual-key"
   

📋 Model Selection Rationale

• Main Agent: Qwen3-235B-A22B-Instruct-2507 - Latest high-performance reasoning model
• Web Agent: Kimi-K2-Instruct - Optimized for web interaction tasks
• Multimodal: Qwen2.5-VL-72B-Instruct - Advanced vision-language capabilities

For all other options, see CONFIG_EXAMPLES.md.

---

📊 GAIA Benchmark Evaluation

Evaluate your agent on the GAIA benchmark:

python -m gaia.cli.simple_validate \
  --data gaia_val.jsonl \
  --level all \
  --count 10 \
  --output results.jsonl

→ Outputs detailed performance summary & per-task results.

---

🌐 Gradio Web UI

Launch a user-friendly web interface:

python -m ck_pro.gradio_app --host 0.0.0.0 --port 7860

→ Open http://localhost:7860 in your browser.

Note: It is recommended to install Playwright browsers (or install them if you encounter related errors). On Linux you may also need to run playwright install-deps.

Note: It is recommended to install Playwright browsers (or install them if you encounter related errors): python -m playwright install (Linux may also require python -m playwright install-deps).

---

📂 Logging

• Console: INFO level by default
• Session logs: logs/ck_session_*.log
• Configurable via [logging] section in TOML

---

🧩 Architecture Highlights

• Modular Design: Web, File, Code, Reasoning modules
• Fallback Mechanism: HTTP API → Playwright browser automation
• Reflection & Voting: Novel test-time strategies for improved accuracy
• Extensible: Easy to plug in new models, tools, or datasets

---

📜 License & Attribution

This is a research-only fork of Tencent’s CognitiveKernel-Pro. 🔗 Original: https://github.com/Tencent/CognitiveKernel-Pro

⚠️ Strictly for academic research and educational purposes. Commercial use is prohibited. See LICENSE.txt for full terms.