AgentCPM-Report Deep Research
Generate comprehensive, long-form research reports using AgentCPM-Report, an open-source 8B parameter agent model that matches top-tier closed-source systems like Gemini-2.5-Pro-DeepResearch. Designed for high-privacy scenarios with fully offline deployment, it performs 40+ rounds of deep retrieval and 100+ rounds of chain-of-thought reasoning to produce logically rigorous, deeply insightful reports with tens of thousands of words.
Capabilities
**Deep Research**: Autonomously conducts extensive research through multi-round retrieval and reasoning**Long-Form Reports**: Generates comprehensive reports with tens of thousands of words**Local & Secure**: Fully offline deployment for complete data privacy**Private Knowledge Base**: Integrates with UltraRAG framework for secure local knowledge retrieval**Competitive Performance**: Matches top closed-source systems with only 8B parametersWhen to Use This Skill
Use this skill when you need to:
Generate comprehensive research reports on complex topicsConduct deep analysis requiring multiple rounds of information gatheringWork with sensitive data that cannot leave your infrastructureProduce detailed, well-structured long-form contentLeverage local private knowledge bases for researchCreate professional decision-making reports from confidential dataInstructions
Follow these steps to generate deep research reports using AgentCPM-Report:
1. Environment Setup
First, set up the AgentCPM-Report environment with Docker:
```bash
git clone https://github.com/OpenBMB/UltraRAG.git
cd UltraRAG
git checkout agentcpm-report-demo
cd agentcpm-report-demo
cp env.example .env
```
2. Deploy the System
Launch the complete stack (UltraRAG, vLLM, Milvus):
```bash
docker-compose -f docker-compose.yml up -d --build
docker-compose -f docker-compose.yml logs -f ultrarag-ui
```
Wait approximately 30 minutes for first startup (image pull, model download, environment configuration). Access the UI at `http://localhost:5050` when ready.
**For CPU-only inference**: Use `docker-compose.cpu.yml` instead for GGUF model support with llama.cpp.
3. Prepare Knowledge Base
Build your research knowledge base:
1. Upload local files through the UI
2. Chunk and index the documents
3. (Optional) Import Wiki2024 corpus as a comprehensive writing database:
- Download from: https://modelscope.cn/datasets/UltraRAG/UltraRAG_Benchmark/tree/master/corpus/wiki24
- Import through the UI
4. Configure Research Task
In the Chat section:
1. Select **AgentCPM-Report** pipeline
2. Ensure the knowledge base is selected as the source
3. Provide clear research instructions with:
- Topic or question to investigate
- Desired scope and depth
- Specific aspects to cover
- Output format preferences
5. Execute Deep Research
Submit your research query and let AgentCPM-Report:
1. Analyze the research question
2. Perform 40+ rounds of deep retrieval from knowledge base
3. Execute 100+ rounds of chain-of-thought reasoning
4. Synthesize information into a comprehensive report
5. Structure findings with logical organization
6. Review and Refine
Once the report is generated:
1. Review the comprehensive output
2. Ask follow-up questions for clarification or expansion
3. Request specific sections to be elaborated
4. Refine the focus or add new research angles
Best Practices
**Knowledge Base Quality**: Use high-quality, relevant documents for better research outcomes**Clear Instructions**: Provide specific research questions and scope to guide the agent**Iterative Refinement**: Start with a broad query, then refine based on initial results**Resource Planning**: Long-form reports may take significant time; plan accordingly**Privacy First**: Leverage local deployment for sensitive research topics**Index Optimization**: Properly chunk and index documents for optimal retrieval performancePerformance Notes
**Model Size**: 8B parameters (efficient for edge deployment)**Benchmark Results**: Competitive with Gemini-2.5-Pro-DeepResearch on DeepResearch Bench, Gym, and DeepConsult**Retrieval Depth**: ~40 rounds of information gathering**Reasoning Steps**: ~100 rounds of chain-of-thought processing**Output Length**: Capable of tens of thousands of words**Resource Requirements**: GPU recommended for optimal performance; CPU inference available via GGUFExample Use Cases
**Enterprise Research**: Analyze market trends using confidential internal data**Academic Literature Review**: Synthesize findings from large corpus of papers**Technical Documentation**: Generate comprehensive technical reports from codebases and specs**Policy Analysis**: Deep dive into policy documents and regulatory materials**Competitive Intelligence**: Research competitors using private intelligence databases**Due Diligence**: Comprehensive analysis for investment or partnership decisionsResources
**Model**: https://huggingface.co/openbmb/AgentCPM-Report**GGUF Version**: https://huggingface.co/openbmb/AgentCPM-Report-GGUF**Framework**: https://github.com/OpenBMB/AgentCPM**UltraRAG**: https://github.com/OpenBMB/UltraRAG**Documentation**: https://ultrarag.openbmb.cn/pages/en/demo/deepresearch**Demo Video**: https://www.youtube.com/watch?v=d5XWONt0PWo**Tutorial Video**: https://www.youtube.com/watch?v=ze8qJRrass4Constraints
Requires Docker and sufficient disk space for model filesFirst-time setup takes ~30 minutesGPU recommended for performance (CPU inference available but slower)Research quality depends on knowledge base comprehensivenessLong-form generation may take extended time for complex topicsDesigned for offline/local deployment (not cloud-optimized)