PlatformIntegrationsMetadata and Smart Filtering

Smart Filtering (beta)

You will often want to do direct filtering along side or instead of semantic search. For example, if a user should be able to ask “show me call id 1234” - you want the Agent to be able to filter down to exactly that call id, rather than any calls that are semantically similar to that. Smart Filtering is our currently recommended way to do that (although note that this feature is still in Beta).

We support smart filtering for all of our data sources. This allows you to filter on fields that you may have imported (from semi-structured data sources like Salesforce and Jira) or metadata you have tagged documents with manually using the data catalog metadata API endpoint, the document upload API, or even automatically extracted metadata using AI extracted entities see this section. This allows us to only send relevant information to the LLM.

For example, you could have a series of call transcripts that you want to upload and tag with which sales rep they correspond to, so that a user can analyze calls just relating to that rep. Once you’ve uploaded the data, you can use the API endpoint above to tag the docs with an “rep_email” field, and then search across only the docs where the rep_email.

To use smart filtering, toggle on the Smart Filtering setting on your agent and make sure you have:

  1. A schema specified for your document collection.
  2. The corresponding metadata fields applied to your documents

This is preconfigured automatically for the Salesforce and Jira data sources - but we recommend that you confirm that the types of these schemas are what you need for your use case.

When an AI agent is given a prompt by a user, the Agent will look at the user prompt and automatically determine which fields should be filtered for based on the user’s prompt. So if the user asks for “calls which ravin@credal.ai was the organizer for”, the agent will check the document collection schema for fields that might allow it to filter down to just those calls.

Note that smart filtering is still in beta. For now, that means it has some limitations, such as the smart filtering step is not a typical action and a separate AI process determines whether or not to apply smart filtering, not the main Agent. This means that your smart filtering decision cannot take advantage of your Agent’s background prompt or data sources in deciding what smart filters to apply. Some users have worked around this by creating an orchestrator agent that determines the smart filters to apply, and then passes that query down to a sub agent to apply the smart filtering. If you have any questions/see any issues, please reach out to the Credal team and we’d be happy to help.

One important note on Smart Filtering: it can take some time between uploading data via the uploadDocumentContents API and the data being available for use in Smart Filtering, even if the uploadDocumentContents API returns successfully. That’s because there is an asynchronous process that embeds the data, which needs to complete for the Smart Filtering functionality to work.