DATAGEN Research Assistant
Automates hypothesis development, data analysis, visualization, and research report production.
The README scopes file activity to a configurable WORKING_DIRECTORY and discloses model providers, search, scraping, GitHub, filesystem MCP, and LangChain tracing integrations. The workflow also includes one human-choice stage, giving partial support for least privilege, confirmation, and data-flow transparency. Deductions apply because permission boundaries for the code agent, filesystem MCP, and tokens are unspecified; confirmation does not clearly gate all later writes or external calls; and tracing destinations are unclear. Keys are shown as placeholders, but there is no guidance on secret storage, log redaction, retention, or third-party privacy. Most dependencies are pinned, while langchain-crw is unpinned and mcp has only a lower bound, with no audit or vulnerability-management evidence. The project warns that data may be modified and recommends backups, but supplies no transactional undo or recovery mechanism. The license identifies an author and external inspiration is linked, yet publisher identity, report-citation guarantees, and maintenance ownership remain incomplete.
The README describes an end-to-end workflow with quality review and revision, and most core packages are version-pinned. The supplied regression tests protect message accumulation and a workflow step limit, demonstrating localized state safeguards. Deductions apply because two narrow tests do not support broad claims of enterprise-grade, consistently reliable, production-ready behavior; the clone URL differs from the assessed repository identity; and the usage instructions contain a numbering-character error. Optional services and reduced capability without Firecrawl are mentioned, but availability handling for browser drivers, model APIs, MCP servers, and loosely constrained dependencies is incomplete. Apart from an ImportError in test-loading code, there is no evidence of user-facing failure messages, retries, graceful degradation, or troubleshooting behavior.
The product identifies scenarios spanning data analysis, hypothesis generation, visualization, search, and report writing, and supports per-agent selection among OpenAI, Google, Anthropic, Ollama, and Groq. CONFIG_DIRECTORY plus skill, tool, and MCP configuration provide useful environmental flexibility. Deductions apply because intended audiences, supported data types, scale limits, and unsuitable scenarios are not defined; boundaries are mostly limited to cost, runtime, and a short issue list. Invocation relies on editing a free-form user_input in main.py. A datapath example exists, but there is no strict input schema, ambiguity handling, or precise task-routing contract. Python, Conda, ChromeDriver, and environment variables are documented, while OS compatibility, containers, and hardware requirements are absent.
The README is organized into overview, features, installation, usage, components, workflow, configuration, issues, contribution, and licensing sections, and it indexes several focused guides. Installation and environment-variable notes are reasonably substantial. Deductions apply because the linked guides are not included in the supplied evidence, the usage section offers only one example, and no FAQ or representative output is shown. The project was renamed, while its clone URL uses starpig1129 and the assessed repository and star-history link use zi-yue-1129, weakening naming and attribution stability. Runtime, possible data mutation, and three current issues are disclosed, earning credit for limitations, though treatment is brief. The complete MIT license merits full credit. There is no version scheme, release policy, or changelog. Contribution guidance exists, but maintainers, support channels, and update responsibility are not clearly assigned, and copyright/repository identities differ.
The workflow combines hypothesis work, analysis code, visualization, search, quality review, and report writing into one configurable process, making the promised graphical report potentially useful for ordinary research tasks. Deductions apply because no actual output, acceptance criteria, or comparative evidence against manual or single-agent workflows is supplied; the claimed marginal benefit of the multi-agent design is largely promotional. The README warns about multiple API calls, credit requirements, and long runtimes, but gives no cost estimate, resource ceiling, caching strategy, or measured return, so cost-benefit support remains thin.
Some implementation-oriented claims trace to the requirements file, configuration example, component list, and two routing regression tests. The README and dependencies also cross-support the stated LangChain, LangGraph, and multi-provider integrations. Deductions apply because major claims such as real-time optimization, automated quality assurance, scalability, and production readiness have no corresponding tests, metrics, sample reports, or implementation excerpts. Cross-source corroboration is limited, and repository/author naming is inconsistent. The documentation does not clearly distinguish implemented behavior from design goals or future vision; the brief current-issues list is insufficient to calibrate its strongest claims.
- Back up data under WORKING_DIRECTORY before use. The README explicitly says the agent may modify analyzed data, and no built-in rollback is documented.
- Use narrowly scoped credentials and an isolated directory for filesystem MCP, code execution, GitHub tokens, and search/scraping services; the supplied material does not define their permission boundaries.
- Do not place API keys or sensitive data in committed .env files, prompts, tracing logs, or generated reports; redaction, retention, and third-party data-handling policies are not documented.
- Pin compatible versions of langchain-crw and mcp and perform dependency vulnerability and supply-chain review before deployment.
- Treat enterprise-grade, production-ready, and quality-assurance claims as insufficiently verified; the supplied tests cover only two state/routing regressions.
What does this agent do, and when should you use it?
DATAGEN is a self-hosted Python platform that coordinates a stateful, multi-agent research workflow with LangGraph. Its named components include hypothesis_agent, process_agent, code_agent, visualization_agent, searcher_agent, report_agent, quality_review_agent, and note_agent. Users place their own dataset in WORKING_DIRECTORY, describe the requested analysis in main.py, and review whether to continue with or regenerate the initial hypothesis. The processing stage combines generated analysis code, visualizations, research searches, report writing, quality review, and revision. It is best suited to researchers or analysts prepared to configure model credentials and manage API usage; adopters should also heed the documented warning that the system may modify input data.
DATAGEN reads a datapath and analysis request from the user_input value in main.py, then accesses the supplied file under WORKING_DIRECTORY. hypothesis_agent generates a research hypothesis and presents a human choice to continue or regenerate it. process_agent supervises the subsequent work performed by code_agent, visualization_agent, searcher_agent, and report_agent, covering analysis-code generation, data analysis, charts, searches, and report production. quality_review_agent reviews the work and initiates revisions when needed, while note_agent records the research process and retains context. LangGraph manages the state graph, and agent_models.yaml lets users assign a provider, model, and temperature to each agent. Optional configuration also covers Filesystem, GitHub, and Web Search MCP servers, along with Firecrawl, fastCRW, Tavily, and LangChain tracing.
- A data analyst has a CSV file and wants to request machine-learning analysis in natural language, with charts and a complete report as outputs.
- A researcher wants an explicit hypothesis-generation checkpoint before committing resources to the full analysis workflow.
- A small research team wants separate agents to handle analysis code, web or literature searches, visualization, report writing, and quality review.
- A team evaluating several model vendors wants to assign OpenAI, Google, Anthropic, Ollama, or Groq models to individual agents.
- An analyst running a long, multi-stage investigation wants note_agent to record progress and preserve context between phases.
What are this agent's strengths and limitations?
- The documented workflow spans hypothesis generation, a human decision point, analysis, visualization, search, report writing, quality review, and revision.
- agent_models.yaml can assign OpenAI, Google, Anthropic, Ollama, or Groq independently to different agents.
- note_agent is dedicated to recording the research process and retaining context, while LangGraph coordinates workflow state.
- The configuration architecture supports a separate CONFIG_DIRECTORY, reusable skills, dynamic tool loading, and optional MCP services.
- Setup requires Python 3.10+, Conda, ChromeDriver, environment variables, and the project's Python dependencies, creating a meaningful self-hosting burden.
- The system makes multiple model API calls; the README warns that sufficient API credits are needed and that complex runs may take time.
- The project warns that agents may modify the analyzed data, so adopters must back up inputs rather than assume read-only processing.
- NoteTaker efficiency, overall runtime, and Refiner quality are explicitly listed as current issues.
- Search functionality may be reduced without optional Firecrawl, fastCRW, Tavily, or MCP credentials and services.
How do you install or deploy this agent?
Python 3.10 or later is required. Run:
git clone https://github.com/starpig1129/DATAGEN.git
cd DATAGEN
conda create -n datagen python=3.10
conda activate datagen
pip install -r requirements.txtRename .env Example to .env. Set WORKING_DIRECTORY, CONDA_ENV, and CHROMEDRIVER_PATH, which the README marks as required. Add OPENAI_API_KEY, ANTHROPIC_API_KEY, or GOOGLE_API_KEY for the provider selected in the agent configuration. CONFIG_DIRECTORY, Firecrawl, fastCRW, LangChain tracing, Tavily, and GitHub MCP settings are documented as optional.
How do you use this agent?
Place a dataset such as YourDataName.csv in the data directory. Edit user_input inside the main() function in main.py:
user_input = '''
datapath:YourDataName.csv
Use machine learning to perform data analysis and write complete graphical reports'''
Start the workflow with:
python main.pyAfter hypothesis generation, choose whether to continue or regenerate the hypothesis. The workflow then performs processing, analysis, visualization, search, report writing, quality review, and any required revision. To change models per agent, edit agent_models.yaml under CONFIG_DIRECTORY and set provider, model, and temperature.