メニュー
Open-source deep research agent by Alibaba for scholarly research
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills --skill tongyi-deep-research-guideコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SOC 職業分類に基づく
SKILL.md を表示中
| name | tongyi-deep-research-guide |
| description | Open-source deep research agent by Alibaba for scholarly research |
| metadata | {"openclaw":{"emoji":"🔎","category":"research","subcategory":"deep-research","keywords":["deep-research","alibaba","tongyi","agentic-rag","scholarly-search","open-source"],"source":"https://github.com/Alibaba-NLP/DeepResearch"}} |
Tongyi DeepResearch is an open-source deep research agent developed by Alibaba's NLP team, with over 18,000 stars on GitHub. It implements an agentic research pipeline that iteratively searches, reads, reasons, and synthesizes information to produce comprehensive research reports. The system is designed to handle complex, multi-faceted research questions that require gathering evidence from multiple sources and reasoning across diverse information.
Unlike simpler RAG (Retrieval-Augmented Generation) systems that perform a single search-and-answer cycle, DeepResearch uses an iterative approach where the agent dynamically decides what to search next based on what it has already found. This makes it particularly effective for research questions that require building up understanding incrementally, following citation chains, or exploring multiple angles of a topic.
The project is notable for being one of the leading open-source alternatives to proprietary deep research tools. It supports multiple LLM backends, various search APIs, and can be customized for domain-specific research needs. For academic researchers, it offers a transparent and modifiable research pipeline where every step can be inspected, reproduced, and adapted.
# Clone the repository
git clone https://github.com/Alibaba-NLP/DeepResearch.git
cd DeepResearch
pip install -r requirements.txt
conda create -n deepresearch python=3.10
conda activate deepresearch
pip install -r requirements.txt
Configure your environment for the LLM and search backends:
# LLM configuration (supports multiple providers)
export LLM_API_KEY=$LLM_API_KEY
export LLM_BASE_URL=$LLM_BASE_URL
export LLM_MODEL=qwen-max
# Search API configuration
export SEARCH_API_KEY=$SEARCH_API_KEY
export SEARCH_ENGINE=bing # or google, serper, tavily
For a fully local deployment with Ollama:
# Use local models
export LLM_BASE_URL=http://localhost:11434/v1
export LLM_MODEL=qwen2.5:72b
export LLM_API_KEY=ollama
DeepResearch follows a think-search-read-reflect loop that mimics how a human researcher works:
from deep_research import DeepResearch
# Initialize the research agent
agent = DeepResearch(
llm_model="qwen-max",
search_engine="bing",
max_iterations=10,
max_sources=30,
)
# Run a research query
result = agent.research(
query="What are the latest advances in multimodal large language models "
"and their applications in scientific research?",
output_format="markdown",
)
print(result.report)
print(f"Sources consulted: {len(result.sources)}")
print(f"Research iterations: {result.iterations}")
Fine-tune the research behavior for different types of queries:
config = {
"max_iterations": 15, # Maximum research cycles
"max_sources_per_query": 10, # Sources per search query
"min_relevance_score": 0.7, # Minimum source relevance threshold
"enable_citation_tracking": True, # Follow citation chains
"language": "en", # Output language
"report_length": "detailed", # brief, standard, or detailed
}
agent = DeepResearch(config=config)
DeepResearch integrates with multiple search providers to cast a wide net:
# Configure multiple search backends for comprehensive coverage
agent = DeepResearch(
search_engines=["bing", "openalex"],
search_strategy="parallel", # Search all engines simultaneously
)
DeepResearch can follow citation chains to discover related work:
result = agent.research(
query="Foundational papers on attention mechanisms in neural networks",
enable_citation_tracking=True,
citation_depth=2, # Follow citations up to 2 levels deep
)
Create research profiles optimized for specific academic domains:
# Biomedical research profile
bio_config = {
"preferred_sources": ["pubmed", "biorxiv", "nature", "science"],
"search_engines": ["openalex", "bing"],
"terminology_mode": "technical",
"citation_format": "apa",
}
agent = DeepResearch(config=bio_config)
result = agent.research(
"Recent developments in mRNA vaccine delivery mechanisms"
)
Monitor the research process in real-time:
async def stream_research():
agent = DeepResearch(llm_model="qwen-max")
async for event in agent.research_stream(
query="Quantum computing applications in drug discovery"
):
if event.type == "thinking":
print(f"Thinking: {event.content}")
elif event.type == "searching":
print(f"Searching: {event.query}")
elif event.type == "reading":
print(f"Reading: {event.url}")
elif event.type == "report":
print(f"Final report:\n{event.content}")
DeepResearch output can be integrated with standard academic tools:
Every research session can be fully reproduced:
# Save the complete research trace
result = agent.research(query="...", save_trace=True)
result.save_trace("research_trace.json")
# Replay a research session
replayed = DeepResearch.replay("research_trace.json")
The trace includes all search queries, retrieved documents, LLM prompts and responses, and reasoning steps, enabling full transparency and reproducibility of the research process.