Get Started

Secoda AI

This page will go over the Secoda AI functionality.

Explore plans which include Secoda AI here.

Overview

The Secoda Secoda AI enhances your data stack by providing a powerful chat interface that allows anyone in your organization to retrieve data insights.

Secoda's Secoda AI is powered by OpenAI's APIs, which sit atop the GPT4 Large Language Model. We are updated to the latest version of GPT -- GPT-4o.

By sharing only the metadata in your workspace with OpenAI, Secoda's AI Assistant can help you with documentation, query building, institutional knowledge discovery, all while safeguarding your metadata!

Enabling the Secoda AI

By default, Secoda AI is disabled. To activate it:

  1. Navigate to Settings > AI.

  2. Toggle on the Secoda AI.

  3. Once enabled, the Secoda AI will appear in the left-hand menu of your workspace.

AI Settings for Admins

Admins can customize AI settings to align with workspace preferences and security protocols:

AI Governance

AI Governance in Secoda empowers organizations to control the data accessed by AI, enhancing security and data relevance.

Key Features:

  • Customizable Filters: Admins can specify which resources are accessible to the AI by setting inclusive or exclusive filter rules. This ensures that AI interactions are confined to appropriate datasets.

  • Consistent Configuration: The filter rules for AI are aligned with those used on the Search and Catalog pages, ensuring a unified approach to data governance across the platform.

Benefits:

  • Enhanced Data Security: Restricts AI access to sensitive or non-production data, minimizing potential exposure.

  • Improved Data Relevance: Focuses AI interactions on current and verified data, increasing the accuracy and reliability of the insights provided.

By implementing these rules, Admins can tailor AI capabilities to fit organizational needs and compliance requirements, ensuring that only relevant and secure data is queried.

Agent Tools

  • By default, tools like search_resources, search_knowledge, catalog_search, retrieve_entity, and entity_link are enabled to ensure comprehensive search capabilities across the workspace.

  • The run_sql tool, disabled by default, allows the AI to execute SQL queries directly on your connected integrations. This tool can provide answers to data queries but can be disabled if there are security concerns. Users can ask questions like "How many customers do we have?" and getting a numeric answer instead of pointing the user to a guiding resource or query. The AI will also provide the user with the steps that it took to find that answer, including any SQL queries it ran on the backend to provide more context.

To ensure security and relevance:

Team-Based Replies: Our AI strictly adheres to Team-based data access permissions. If a user asks about resources from a Team that they are not a part of, the AI will not return any results.

Query Execution Requirements: The AI will only execute a query if both of the following conditions are met:

  1. run_sql is toggled on in the AI settings.

  2. The user asking the question has Query access for the specific integration, which can be managed by admins in the Integration Permissions settings.

This ensures that only authorized users can run queries on sensitive or critical data, maintaining control over data access.

Custom Instructions

Admins can set specific instructions for the AI to follow, enhancing control and relevance.

Some examples that we've seen work well:

  • Do not index on resources tagged with "archived" or "deprecated" or "stale".

  • Only provide results on "Verified" resources.

  • Only provide results on Published resources - do not include resources that are still in Draft.

  • Only reference tables and views from the production analytics schema in Snowflake, never use data from the RAW database.

By default, both Published and Draft resources are included in Secoda AI responses. This can be controlled using the custom instructions to restrict this to only Published resources, for example.

Descriptions custom instructions

Admins are able to define custom instructions to the AI Description Editor so that you can define documentation standards and a format for it to follow.

Some examples that we've seen work well:

  • Add "AI-Generated" to the end of each description added, to indicate to users that a human didn't come up with this definition.

  • Do not reference the database, schema, table information in the description (this information is obvious in Secoda)

  • Do not use full, wordy sentences. Be brief and include only the necessary facts.

  • Descriptions should be concise and not editorialized. Do not describe the data as 'essential' or 'critical'; only describe the contents.

  • All tables and columns are related to insurance. Keep this in mind when generating descriptions.

Future Improvements

Looking ahead, we are committed to enhancing the Secoda AI Assistant with:

  • Improved response times and reliability.

  • Enhanced understanding of complex prompts.

  • Tailored responses based on user roles and access levels.

Additional documentation

Secoda AI User GuideSecoda AI Use CasesAI Security FAQsPrompts

This guide ensures that Admins are equipped to optimize the AI Assistant within their Secoda environment, driving efficient and secure data operations across the organization.

Secoda AI Assistant is only available for Version 7.0.0 and up.

Last updated