data-access-and-eda

# data-access-and-eda Use this sub-skill when the task is about discovering local data files, loading tabular files, summarizing DataFrames, or using the package's tool-calling data loader / EDA agents. ## Best-fit tasks - List, search, or inspect files and folders before analysis. - Load CSV, TSV, Excel, JSON/JSONL/NDJSON, Parquet, or explicitly trusted pickle files into pandas. - Summarize one or more pandas DataFrames without asking an LLM to write pandas code. - Use `DataLoaderToolsAgent` to let a chat model choose data-loader tools. - Use `EDAToolsAgent` for explain/describe/missingness/correlation/report tool selection. - Decide whether optional EDA report tooling is appropriate for a safe, bounded run. ## Route away from this sub-skill - Cleaning, wrangling, feature engineering, or visualization code generation: use `dataframe-code-agents`. - SQL database metadata, query generation, or SQL safety: use `sql-analysis`. - Streamlit apps, Pipeline Studio, supervisor workflows, or multi-agent orchestration: use `multiagent-and-app-workflows`. - H2O AutoML, model evaluation, MLflow tools, or ML service troubleshooting: use `modeling-and-mlflow`. ## Fast operating path 1. For deterministic data loading and summaries, prefer direct tool functions and DataFrame helpers from [references/api-reference.md](references/api-reference.md). 2. For user-facing file discovery or EDA requests that genuinely need model-based tool selection, use the agent workflows in [references/workflows.md](references/workflows.md). These require a LangChain-compatible model and may call that model when invoked. 3. Treat optional EDA report tools as opt-in. Check [references/optional-eda-reports.md](references/optional-eda-reports.md) before using `missingno`, `pytimetk`, `sweetviz`, or `dtale` features. 4. If a symptom appears, use [references/troubleshooting.md](references/troubleshooting.md) before retrying. 5. For a no-network, no-LLM smoke check, run [`scripts/smoke_data_access.py`](scripts/smoke_data_access.py) in the target Python environment. ## Safety defaults - Do not call external LLM providers unless the user explicitly provides a model configuration and expects agent invocation. - Use listing/search tools before loading file contents when the user only asks what files exist. - Do not load pickle files from untrusted sources. Pickle loading is disabled by default and requires an explicit `ALLOW_UNSAFE_PICKLE` opt-in in the process environment. - Use small sample sizes for summaries of wide or large DataFrames. - Optional report tools may write HTML files or launch local services; get explicit approval before browser, service, or long-running report workflows.