Skip to content

What Depositori reads and sends

Depositori keeps every model it reads in Apache Ossie, an open format for semantic models: tables, columns, relationships and metrics, written the same way whatever tool they came from. That is why a dbt model, a Power BI model and a Cube model look alike in Depositori, and why their lineage joins up.

Most sources are read into Ossie by Apache Ossie’s own converters, one per tool, so Depositori reads those tools the way the Ossie project does. Amazon QuickSight and Apache Superset have no Ossie converter: Depositori converts them itself. A few converters are not in Depositori yet (Polaris, RelationalAI, Ataccama); the rest are.

Under Add a model, in the same four groups:

Group Sources
BI tools Power BI (a workspace, or a Power BI template file), Amazon QuickSight, Apache Superset, Sigma, Omni, ThoughtSpot, GoodData, Holistics, Lightdash, Salesforce
Semantic layers dbt, Cube, Honeydew, OrionBelt
Data platforms Databricks (a Unity Catalog metric view); and through a connection, for the lineage: Azure SQL, PostgreSQL, Snowflake, DuckLake
AI platforms WisdomAI, NVIDIA Auto Ontology, Palantir Foundry (an ontology export: its object types’ datasets, keys and links)

Most of them are read from the files the tool exports, with no account at the source. Power BI, dbt, the databases, Superset and QuickSight can also be read through a connection. A source that can’t be read on your Depositori says Not available here.

The other way round, Compare writes a model out in the format of another tool, again with Apache Ossie’s converters: most of the tools above (not QuickSight, Superset or Palantir Foundry), and also Looker (LookML for BigQuery), Hex and Snowflake. Nothing is written back to a source: you download the files.

Depositori never changes a source. It reads definitions: tables, columns, relationships, metrics, views and their SQL, datasets, descriptions.

It reads data only to answer a question:

  • from a Power BI model, for Taki’s live numbers, as the person asking or the organisation’s app account;
  • from a database whose connection has Taki may query on;
  • from a tool server (MCP) behind Depositori, when an agent calls one of its read-only tools;
  • for an application with an API key, when an Admin turned on Allow queries through the API.

To pick the right value for a question (a region, a product), Taki may also look up the values of a column: at most 200 of them.

It reads a source as the source account the connection names, so it sees no more than that account may see.

  • The definitions it read, as models.
  • Your organisation’s glossary, notes, answers to open questions, and the descriptions written by hand.
  • Source accounts: passwords, secrets, keys and tokens are stored encrypted, and nobody can view them again.

Only with AI on, for the organisation (see Taki, the AI analyst):

With AI on, the model definitions and glossary of your organisation, your questions, and what Taki’s queries return, go to Anthropic.

What Taki’s queries return is:

  • for a Power BI model: the answers Power BI gives to Taki’s questions, read with the account of the person asking or the organisation’s app account;
  • for a database with Taki may query on: the rows of its read-only queries. It skips columns whose names point to personal data, such as names, email addresses and phone numbers, health and absence (sick leave, verzuim) and personnel numbers and keys. A score or check of something else, such as a customer account’s health_score, stays readable. Taki goes by names only: for HR and absence data, give the connection’s login access to aggregated views only. Only the connection’s owner turns this on;
  • for a tool server (MCP) behind Depositori: what its read-only tools return;
  • the values of a column it looked up, at most 200.

With AI off, nothing goes to Anthropic.

It goes there with your own key, the organisation’s or the workspace’s: Depositori has no AI key of its own. So your own agreement with Anthropic applies to it, not one of In2Intel’s.

When Depositori runs in the cloud, it guards your network and your accounts. You may meet these rules:

  • Public addresses only. A connection to an internal address (a private network or the cloud’s own services) is refused, with “… is not an address this app may connect to”. Use the source’s public host name.
  • Tool servers over https. An MCP server’s address must start with https://.
  • Superset over https. A Superset address must be its public https:// address.
  • A DuckLake’s data on Azure Storage, in the storage account named on the connection; its catalog in PostgreSQL. Your Microsoft account’s access goes only to that storage account. See DuckLake.
  • Taki queries tables only, never files or web addresses.
  • Errors as categories. A failed test or read says what kind of problem it is, such as “The server’s firewall doesn’t allow this app.” or “The server didn’t answer.”, with what to do. The database driver’s own message is not shown: it can hold addresses and user names.
  • AWS keys of their own. A QuickSight connection reads only with the AWS access key you stored, never with the identity of the machine Depositori runs on. Temporary keys are refused.