Analyze Local Tabular Files

#Goal

Analyze local exports without attaching a directory as a database source or exposing arbitrary filesystem paths.

Suppose the admitted workspace contains:

/absolute/path/project/
└── exports/
    ├── orders.csv
    ├── customers.ndjson
    └── archive/
        └── prior-orders.csv

#Admit the Workspace

from pathlib import Path
 
from daita import Agent, LocalWorkspace
 
workspace = LocalWorkspace(Path("/absolute/path/project"))
 
async with await Agent.open("exports", workspace=workspace) as agent:
    result = await agent.run(
        "Query exports/orders.csv. Summarize order count by status and "
        "identify how many rows have a missing customer_id.",
        files_only=True,
    )
    print(result.final_text)

The workspace is a runtime admission, not a SourceRegistration. It has no source_id, does not create catalog resources, and does not require refresh after a file changes.

#What the Agent Can Use

The Files toolbox provides:

• file_search to locate files by workspace-relative path or literal text;
• file_read to read one bounded text chunk; and
• file_query to filter or aggregate a homogeneous CSV, TSV, JSON-records/NDJSON, or Parquet dataset.

For the aggregation above, file_query expands an exact relative pattern and exposes the selected dataset to validated SQL as one relation named data. Daita records the complete input manifest and physical revisions, runs a private one-call DuckDB worker, and returns at most the bounded result projection.

#Containment

The model receives workspace-relative logical paths, never the absolute root. Daita rejects traversal, symlinks, special files, secret-like paths, mixed or incompatible datasets, and files that change during a cursor or query binding.

There is no attached CSV/JSON source and no data_read_file tool in the current framework.

See Local Workspace for file-query bounds and the approved text-edit workflow.