Analyze Local CSV and JSON Files

#Goal

Make a directory of CSV and JSON exports available to an agent without exposing arbitrary filesystem paths.

Suppose the directory contains:

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

#Attach the Directory

from daita import Agent, LocalDirectorySource
 
async with await Agent.open("exports") as agent:
    source = await agent.attach(
        LocalDirectorySource(
            "/absolute/path/exports",
            name="Daily exports",
            max_depth=3,
            max_files=100,
            max_file_bytes=2 * 1024 * 1024,
            max_rows=100_000,
        )
    )
 
    resources = await agent.list_catalog_resources(source_id=source.id)
    for resource in resources:
        print(resource.id, resource.name, resource.revision)
 
    result = await agent.run(
        "Summarize orders.csv by status and identify missing customer IDs.",
        source_id=source.id,
    )
    print(result.final_text)

Only .csv and .json resources within the admitted root are discovered. Symlinks, path escapes, non-regular files, and files outside the configured bounds are rejected.

#Preview the Catalog

The catalog loop in the example shows the stable resource identity used by the model-facing read tool. The model never supplies an arbitrary filesystem path.

#Analyze a File

The final run() lets the agent use data_read_file, which returns a bounded row projection with truncation and revision facts. The agent should describe limits when it has not seen the complete resource.

#Refresh the Snapshot

After adding or replacing files:

Call await agent.refresh_source(source.id) while the agent is open.

Daita checks containment and file identity again at read time. A replaced or changed file is rejected until refresh commits its new catalog revision.

See CSV and JSON for discovery and containment details.