Skip to content

Connect data

You can ask the agent to connect a supported data source for you, or connect it directly with /connect. Direct connections are useful when you want to choose the alias yourself, start browsing immediately, or run TabulaFlow with the LLM turned off. Once connected, the source appears in the Data Explorer, where you can browse its schemas, tables, and data without an active LLM.

For example, ask the agent:

Connect ./data/orders.parquet as orders.

Connect directly

Run /connect in TabulaFlow:

/connect <source...> [--alias name]

Connections last for the current session. TabulaFlow creates an alias for each source; use --alias to choose your own.

Choose your source

You have Connect with Details
CSV, Excel, Parquet, or JSON files One or more file paths Local files
SQLite or DuckDB database A database file path Local databases
SQL database supported by SQLAlchemy A connection URL SQL databases
Neo4j graph database A neo4j or bolt URL Neo4j
SPARQL endpoint A sparql+http or sparql+https URL SPARQL
Hugging Face dataset A dataset or viewer URL Hugging Face
Wikidata knowledge graph The name wikidata Wikidata

Local files

You can connect CSV, TSV, Excel (.xlsx and .xls), Parquet, JSON, JSONL, and NDJSON files.

/connect ./data/orders.parquet

Relative paths start from the directory where you launched TabulaFlow. Each file is imported into a session-owned DuckDB table named after the file. Connect related files together to group them under one source:

/connect ./data/customers.csv ./data/orders.csv --alias retail

Local databases

Connect SQLite (.sqlite, .sqlite3, or .db) and DuckDB (.duckdb) files by path:

/connect ./data/analytics.duckdb --alias analytics

SQL databases

TabulaFlow accepts SQLAlchemy-compatible database URLs. Drivers for PostgreSQL, MySQL, Snowflake, and BigQuery are included; other SQLAlchemy dialects can be used when their required driver packages are installed.

Source Example
PostgreSQL postgresql://user@localhost/analytics
MySQL mysql://user@localhost/analytics
Snowflake snowflake://user@account/database
BigQuery bigquery://project/dataset
/connect postgresql://user@localhost/analytics --alias warehouse

TabulaFlow automatically selects async drivers for PostgreSQL and MySQL.

Neo4j

Connect a Neo4j database with the URL provided by your deployment:

/connect neo4j+s://user@host?database=neo4j --alias graph

Use the connection URI supplied by your Neo4j deployment unchanged, including its scheme (neo4j, neo4j+s, bolt, or bolt+s). Do not replace it with the scheme shown in this example: the scheme determines routing, TLS, and certificate verification.

SPARQL

Prefix the endpoint URL with sparql+:

/connect sparql+https://example.org/sparql --alias knowledge_graph

TabulaFlow does not infer that a plain HTTP URL is a SPARQL endpoint.

Hugging Face

Connect a dataset with its Hugging Face URL:

/connect https://huggingface.co/datasets/nyu-mll/glue/viewer/sst2/train

If a dataset has multiple configurations, TabulaFlow asks you to choose one. Add /viewer/<subset>/<split> to the URL to select both directly.

Wikidata

Connect the Wikidata knowledge graph by name:

/connect wikidata

TabulaFlow connects to the Wikidata Query Service as a SPARQL source and provides source-specific query guidance to the agent.

Source safety and the workspace

Connected sources are read-only, so browsing and analysis do not change the original data. The agent uses a local DuckDB workspace as its writable area, storing anything it needs to create or retain for the session, including intermediate results, transformed or combined data, and extracted records. You can ask the agent at any time to export results to local files in any format you need for saving, sharing, or further use.

Credentials

Use environment variables, cloud-provider configuration, or your driver's credential store. Keep passwords out of shell history, committed files, prompts, and issue reports. If a URL must contain credentials, percent-encode reserved characters and use a least-privilege read-only account.

TabulaFlow may send content needed for a prompt to your model provider. Review your provider's data-handling and retention policies before connecting sensitive data.

Disconnect a source

/disconnect warehouse