Knowledge assist filters

Filters help you narrow down a search and ensure your knowledge assist generator produces useful suggestions. Filter documents in a Datastore to specify a category of documents for reference in both generative knowledge assist (GKA) and proactive generative knowledge assist (PGKA). Customize safety filters on generated content to ensure PGKA suggestions are maximally helpful and minimally harmful for your business needs.

Document filters

Filter knowledge assist documents for both GKA and PGKA with SearchConfig.

Step 1: Set up Datastore with metadata

Import with metadata to create a Datastore for GKA and PGKA. You must provide metadata as one or more JSONL files, located in the higher-level folder of the content used for import.

1. Follow the steps on the Datastore tools page to add a Datastore to an agent.
2. On the Vertex AI Search page, click Data Stores > + Create Data Store.
3. Navigate to Cloud Storage, and click Select.
4. Click File, then enter the path to your file in Cloud Storage or click Browse to select your file.
5. Navigate to What kind of data are you importing? and select Linked unstructured documents (JSONL with metadata).
6. Navigate to Data import options and select Full.
7. Click Import.
8. After you successfully create your Datastore, view the Documents table. The table lists each of your documents with the following details.
   • ID
   • URI link
   • Index status
   • Actions
9. Click View Document to examine a document's properties.

The following is an example JSONL file. You can also specify metadata in the structData field.

{ "id": "d001", "content": {"mimeType": "text/html", "uri": "gs://example-import/unstructured/first_doc.html"}, "structData": {"property0": "value0", "property1": "value1", "property2": "value2"} }
{ "id": "d002", "content": {"mimeType": "application/pdf", "uri": "gs://example-import/unstructured/second_doc.pdf"}, "structData": {"property0": "value0", "property1": "value3", "property2": "value4"} }

Step 2: Configure filters

To directly apply filter values to a GKA search, call the SearchKnowledge API. You can also apply a SearchConfig field in the request, as follows. See the Vertex AI Search documentation for more examples of filter expressions.

{
  "query": {
     "text": "test query"
  },
  "conversationProfile": "projects/PROJECT_ID/locations/LOCATION_ID/conversationProfiles/CONVERSATION_PROFILE_ID",
  "sessionId": "SESSION_ID",
  "conversation": "projects/PROJECT_ID/locations/LOCATION_ID/conversations/CONVERSATION_ID",
  "querySource": "AGENT_QUERY",
  "searchConfig": 
    "filterSpecs": [{
      "dataStores": ["projects/PROJECT_NUMBER/locations/LOCATION_ID/collections/default_collection/dataStores/DATASTORE_ID"],
      "filter": "property1 : ANY(\"value1\")" 
      }]
    }
  }
}

To apply filter values across all knowledge search queries and suggestions within a single conversation, use IngestContextReferences and attach the conversation ID to the SearchKnowledge call.

PGKA document filters

Apply a SearchConfig to PGKA calls using either the IngestContextReferences API or the simulator in the Agent Assist console.

{
  "conversation": "projects/PROJECT_ID/locations/global/conversations/CONVERSATION_ID",
  "contextReferences": {
   "gka_search_config": {
      "contextContents": [{
"content": "{ \"filterSpecs\": [{ \"dataStores\": [\"projects/PROJECT_NUMBER/locations/LOCATION_ID/collections/default_collection/dataStores/DATASTORE_ID\"], \"filter\": \"property1: ANY    (\\\"value1\\\")\" }] }",
        "contentFormat": "JSON"
      }],
      "updateMode": "OVERWRITE",
      "languageCode": "en-US"
    }
  }
}

{ "filterSpecs": [{"dataStores" : ["projects/PROJECT_NUMBER/locations/LOCATION_ID/collections/default_collection/dataStores/DATASTORE_ID"], "filter" : "property1 : ANY(\"value1\")" }] }

Safety filters

Generative features in Agent Assist have built-in responsible AI (RAI) safety filters. These filters are essential for preventing the generation of potentially harmful or inappropriate content. However, default safety levels can be too restrictive for certain industries. For example, a healthcare provider might need to respond to legitimate queries about sensitive health topics that default filters block. To align the generator's behavior with your business needs and use cases, you can customize the sensitivity of safety filters for the PGKA feature.

Basics

There are two key components to safety filters:

• The category of content for filtering
• The sensitivity level of the filter

RAI categories

You can set a sensitivity level for the following content categories:

• Dangerous content: Content that refers to self harm or anything that is illegal, regulated, or dangerous.
• Sexually explicit: Content that is sexually explicit or contains pornographic material.
• Harassment: Content that is insulting, intimidating, or abusive.
• Hate speech: Content that promotes violence or incites hatred against individuals or groups based on certain attributes.

Sensitivity levels

For each RAI category, you can choose one of the following sensitivity thresholds:

• BLOCK_MOST: Blocks a wide range of content that could potentially fall into the category.
• BLOCK_SOME: Blocks content that is clearly identified as belonging to the category.
• BLOCK_FEW: Blocks only the most severe instances of content for the category.
• BLOCK_NONE: Disables all filters for the specified category.

Default settings

If you don't provide a custom configuration for a category, proactive generative knowledge assist uses the following default sensitivity levels:

• Dangerous content: BLOCK_FEW
• Sexually explicit: BLOCK_SOME
• Harassment: BLOCK_SOME
• Hate speech: BLOCK_SOME

Configure safety filters

Configure safety filters within your conversation profile. You can specify a sensitivity level for one or more RAI categories by adding an rai_settings object to the feature configuration for KNOWLEDGE_ASSIST. You can find safety filter settings within the feature_configs array for the KNOWLEDGE_ASSIST suggestion feature.

{
  "name": "projects/PROJECT_ID/locations/LOCATION/conversationProfiles/PROFILE_ID",
  "human_agent_assistant_config": {
    "human_agent_suggestion_config": {
      "feature_configs": [
        {
          "suggestion_feature": {
            "type": "KNOWLEDGE_ASSIST"
          },
          "rai_settings": {
            "rai_category_configs": [
              {
                "category": "DANGEROUS_CONTENT",
                "sensitivity_level": "BLOCK_FEW"
              },
              {
                "category": "HARASSMENT",
                "sensitivity_level": "BLOCK_MOST"
              }
            ]
          }
        }
      ]
    }
  }
}

Configuration examples

The following examples illustrate how to increase or decrease the sensitivity level for a single category according to your business needs.

Example 1: Allow sensitive healthcare queries

As a healthcare provider, you want to ensure that queries about sensitive mental health topics are not blocked. You can lower the sensitivity for the dangerous content category, as follows.

"rai_settings": {
  "rai_category_configs": [
    {
      "category": "DANGEROUS_CONTENT",
      "sensitivity_level": "BLOCK_NONE"
    }
  ]
}

With this setting, proactive generative knowledge assist is more likely to process and provide a helpful response for a query like What are the warning signs of suicide?.

Example 2: Increase strictness

If your business wants to be extra cautious about harassment, you can increase the sensitivity for that category while leaving others at their default levels, as follows.

"rai_settings": {
  "rai_category_configs": [
    {
      "category": "HARASSMENT",
      "sensitivity_level": "BLOCK_MOST"
    }
  ]
}