Filter Configuration

This page describes how to enable filtering features on the Vald cluster.

Before using the filtering functions, please check the Vald Filter Gateway document first for what you can do.

Requirement

To use any filtering functions with the Vald cluster, you must deploy the ingress and/or egress filter component before deploying the Vald cluster.

The ingress filter can be used for, e.g., converting the object data to the vector, some filtering query vector, as pre-processing.

The egress filter can be used for, e.g., filtering search result from vald-lb-gateway by the distance, categories, or any other condition, as post-processing.

The filter component can be deployed anywhere, but it must be able to communicate with the Vald Filter gateway. Every filter component should meet Vald’s filter gRPC interface.

• The ingress RPC definition

  service Filter {
    // Represent the RPC to generate the vector.
    rpc GenVector(payload.v1.Object.Blob) returns (payload.v1.Object.Vector) {
      option (google.api.http) = {
        post : "/filter/ingress/object"
        body : "*"
      };
    }
    // Represent the RPC to filter the vector.
    rpc FilterVector(payload.v1.Object.Vector)
        returns (payload.v1.Object.Vector) {
      option (google.api.http) = {
        post : "/filter/ingress/vector"
        body : "*"
      };
    }
  }
  

• The egress RPC definition

  service Filter {
    // Represent the RPC to filter the distance.
    rpc FilterDistance(payload.v1.Object.Distance)
        returns (payload.v1.Object.Distance) {
      option (google.api.http) = {
        post : "/filter/egress/distance"
        body : "*"
      };
    }
    // Represent the RPC to filter the vector.
    rpc FilterVector(payload.v1.Object.Vector)
        returns (payload.v1.Object.Vector) {
      option (google.api.http) = {
        post : "/filter/egress/vector"
        body : "*"
      };
    }
  }
  

For more details, please refer to the Vald Filter Gateway document

Official filter components

Vald provides the two types of ingress filter components. These will help you to implement your original filter component.

Please refer to:

• Vald ONNX Ingress Filter
• Vald Tensorflow Ingress Filter

Configuration

It is easy to enable the filtering feature.

---
gateway:
---
filter:
  enabled: true

The Vald Filter gateway connects to the filter component specified in the users’ request. So, you can use the filtering function only by setting gateway.filter.enabled=true in your Helm chart.

If you want to make more detailed settings, please set the following parameters.

gateway:
  filter:
  ...
    gateway_config:
    ...
      # gRPC client config for ingress filter
      ingress_filter:
        # gRPC client for ingress filter (overrides defaults.grpc.client)
        client: {}
        # object ingress vectorize filter targets
        vectorizer: ""
        # search ingress vector filter targets
        search_filters: []
        # insert ingress vector filter targets
        insert_filters: []
        # update ingress vector filter targets
        update_filters: []
        # upsert ingress vector filter targets
        upsert_filters: []
      # gRPC client config for egress filter
      egress_filter:
        # gRPC client config for egress filter (overrides defaults.grpc.client)
        client: {}
        # object egress vector filter targets
        object_filters: []
        # distance egress vector filter targets
        distance_filters: []

Those parameters help the Vald filter gateway connect the filter component before getting the users’ requests.

Client configuration

To use the filter function, you must make settings on the client side because the Vald Filter gateway does not automatically create a connection with the filter component.

These sample code describes how to use in client-side when you send the request with the blob data and want to convert it to a vector using the filter component.

package main

import (
	"context"

	"github.com/vdaas/vald-client-go/v1/payload"
	"github.com/vdaas/vald-client-go/v1/vald"
	"github.com/vdaas/vald/internal/log"
	"google.golang.org/grpc"
)

func main() {
	// address of the Vald cluster ingress.
	grpcServerAddr := "vald-ingress-host"

	// ingress filter host and port.
	const ingressHost = "vald-onnx-ingress-filter"
	const ingressPort = 8081

	// object data and its id.
	const id = "object-id"
	var object []byte

	// connect to the Vald cluster
	ctx := context.Background()
	conn, err := grpc.NewClient(grpcServerAddr, grpc.WithInsecure())
	if err != nil {
		log.Error(err)
		return
	}

	// create client
	client := vald.NewFilterClient(conn)
	// config for insert request
	icfg := &payload.Insert_ObjectRequest{
		// object data you'd like to insert into the Vald cluster
		Object: &payload.Object_Blob{
			Id: id,
			Object: object,
		},
		// insert config
		Config: &payload.Insert_Config{
			SkipStrictExistCheck: false,
		},
		// specify vectorizer component location
		Vectorizer: &payload.Filter_Target{
			Host: ingressHost,
			Port: ingressPort,
		},
	}

	// send insertObject request
	res, err := client.InsertObject(ctx, icfg)
	if err != nil {
		log.Error(err)
		return
	}
	log.Infof("location: %#v", res.Ips)
}

If you are the Vald cluster operator and not a client user, please share this information with client users.

See also