AI Proxy Advanced Configuration

You are browsing unreleased documentation.

Configuration

This plugin is compatible with DB-less mode.

Compatible protocols

The AI Proxy Advanced plugin is compatible with the following protocols:

grpc, grpcs, http, https

Parameters

Here's a list of all the parameters which can be used in this plugin's configuration:

• name or plugin

  string required

  The name of the plugin, in this case ai-proxy-advanced.

  • If using the Kong Admin API, Konnect API, declarative configuration, or decK files, the field is name.
  • If using the KongPlugin object in Kubernetes, the field is plugin.

• instance_name

  string

  An optional custom name to identify an instance of the plugin, for example ai-proxy-advanced_my-service.

  The instance name shows up in Kong Manager and in Konnect, so it's useful when running the same plugin in multiple contexts, for example, on multiple services. You can also use it to access a specific plugin instance via the Kong Admin API.

  An instance name must be unique within the following context:

  • Within a workspace for Kong Gateway Enterprise
  • Within a control plane or control plane group for Konnect
  • Globally for Kong Gateway (OSS)

• service.name or service.id

  string

  The name or ID of the service the plugin targets. Set one of these parameters if adding the plugin to a service through the top-level /plugins endpoint. Not required if using /services/{serviceName|Id}/plugins.

• route.name or route.id

  string

  The name or ID of the route the plugin targets. Set one of these parameters if adding the plugin to a route through the top-level /plugins endpoint. Not required if using /routes/{routeName|Id}/plugins.

• consumer.name or consumer.id

  string

  The name or ID of the consumer the plugin targets. Set one of these parameters if adding the plugin to a consumer through the top-level /plugins endpoint. Not required if using /consumers/{consumerName|Id}/plugins.

• consumer_group.name or consumer_group.id

  string

  The name or ID of the consumer group the plugin targets. If set, the plugin will activate only for requests where the specified group has been authenticated /plugins endpoint. Not required if using /consumer_groups/{consumerGroupName|Id}/plugins.

• enabled

  boolean default: true

  Whether this plugin will be applied.

• config

  record required
  • balancer

    record required
    • algorithm

      string default: round-robin Must be one of: round-robin, lowest-latency, lowest-usage, consistent-hashing, semantic

      Which load balancing algorithm to use.

    • tokens_count_strategy

      string default: total-tokens Must be one of: total-tokens, prompt-tokens, completion-tokens

      What tokens to use for usage calculation. Available values are: total_tokens prompt_tokens, and completion_tokens.

    • latency_strategy

      string default: tpot Must be one of: tpot, e2e

      What metrics to use for latency. Available values are: tpot (time-per-output-token) and e2e.

    • hash_on_header

      string default: X-Kong-LLM-Request-ID

      The header to use for consistent-hashing.

    • slots

      integer default: 10000 between: 10 65536

      The number of slots in the load balancer algorithm.

    • retries

      integer default: 5 between: 0 32767

      The number of retries to execute upon failure to proxy.

    • connect_timeout

      integer default: 60000 between: 1 2147483646

    • write_timeout

      integer default: 60000 between: 1 2147483646

    • read_timeout

      integer default: 60000 between: 1 2147483646

  • embeddings

    record
    • auth

      record
      • header_name

        string referenceable

        If AI model requires authentication via Authorization or API key header, specify its name here.

      • header_value

        string referenceable encrypted

        Specify the full auth header value for ‘header_name’, for example ‘Bearer key’ or just ‘key’.

      • param_name

        string referenceable

        If AI model requires authentication via query parameter, specify its name here.

      • param_value

        string referenceable encrypted

        Specify the full parameter value for ‘param_name’.

      • param_location

        string Must be one of: query, body

        Specify whether the ‘param_name’ and ‘param_value’ options go in a query string, or the POST form/JSON body.

      • azure_use_managed_identity

        boolean default: false

        Set true to use the Azure Cloud Managed Identity (or user-assigned identity) to authenticate with Azure-provider models.

      • azure_client_id

        string referenceable

        If azure_use_managed_identity is set to true, and you need to use a different user-assigned identity for this LLM instance, set the client ID.

      • azure_client_secret

        string referenceable encrypted

        If azure_use_managed_identity is set to true, and you need to use a different user-assigned identity for this LLM instance, set the client secret.

      • azure_tenant_id

        string referenceable

        If azure_use_managed_identity is set to true, and you need to use a different user-assigned identity for this LLM instance, set the tenant ID.

      • gcp_use_service_account

        boolean default: false

        Use service account auth for GCP-based providers and models.

      • gcp_service_account_json

        string referenceable encrypted

        Set this field to the full JSON of the GCP service account to authenticate, if required. If null (and gcp_use_service_account is true), Kong will attempt to read from environment variable GCP_SERVICE_ACCOUNT.

      • aws_access_key_id

        string referenceable encrypted

        Set this if you are using an AWS provider (Bedrock) and you are authenticating using static IAM User credentials. Setting this will override the AWS_ACCESS_KEY_ID environment variable for this plugin instance.

      • aws_secret_access_key

        string referenceable encrypted

        Set this if you are using an AWS provider (Bedrock) and you are authenticating using static IAM User credentials. Setting this will override the AWS_SECRET_ACCESS_KEY environment variable for this plugin instance.

      • allow_override

        boolean default: false

        If enabled, the authorization header or parameter can be overridden in the request by the value configured in the plugin.

    • model

      record required
      • provider

        string required Must be one of: openai, mistral

        AI provider format to use for embeddings API

      • name

        string required

        Model name to execute.

      • options

        record

        Key/value settings for the model

        • upstream_url

          string

          upstream url for the embeddings

  • vectordb

    record
    • strategy

      string required Must be one of: redis

      which vector database driver to use

    • dimensions

      integer required

      the desired dimensionality for the vectors

    • threshold

      number required

      the default similarity threshold for accepting semantic search results (float)

    • distance_metric

      string required Must be one of: cosine, euclidean

      the distance metric to use for vector searches

    • redis

      record required
      • host

        string default: 127.0.0.1

        A string representing a host name, such as example.com.

      • port

        integer default: 6379 between: 0 65535

        An integer representing a port number between 0 and 65535, inclusive.

      • connect_timeout

        integer default: 2000 between: 0 2147483646

        An integer representing a timeout in milliseconds. Must be between 0 and 2^31-2.

      • send_timeout

        integer default: 2000 between: 0 2147483646

        An integer representing a timeout in milliseconds. Must be between 0 and 2^31-2.

      • read_timeout

        integer default: 2000 between: 0 2147483646

        An integer representing a timeout in milliseconds. Must be between 0 and 2^31-2.

      • username

        string referenceable

        Username to use for Redis connections. If undefined, ACL authentication won’t be performed. This requires Redis v6.0.0+. To be compatible with Redis v5.x.y, you can set it to default.

      • password

        string referenceable encrypted

        Password to use for Redis connections. If undefined, no AUTH commands are sent to Redis.

      • sentinel_username

        string referenceable

        Sentinel username to authenticate with a Redis Sentinel instance. If undefined, ACL authentication won’t be performed. This requires Redis v6.2.0+.

      • sentinel_password

        string referenceable encrypted

        Sentinel password to authenticate with a Redis Sentinel instance. If undefined, no AUTH commands are sent to Redis Sentinels.

      • database

        integer default: 0

        Database to use for the Redis connection when using the redis strategy

      • keepalive_pool_size

        integer default: 256 between: 1 2147483646

        The size limit for every cosocket connection pool associated with every remote server, per worker process. If neither keepalive_pool_size nor keepalive_backlog is specified, no pool is created. If keepalive_pool_size isn’t specified but keepalive_backlog is specified, then the pool uses the default value. Try to increase (e.g. 512) this value if latency is high or throughput is low.

      • keepalive_backlog

        integer between: 0 2147483646

        Limits the total number of opened connections for a pool. If the connection pool is full, connection queues above the limit go into the backlog queue. If the backlog queue is full, subsequent connect operations fail and return nil. Queued operations (subject to set timeouts) resume once the number of connections in the pool is less than keepalive_pool_size. If latency is high or throughput is low, try increasing this value. Empirically, this value is larger than keepalive_pool_size.

      • sentinel_master

        string

        Sentinel master to use for Redis connections. Defining this value implies using Redis Sentinel.

      • sentinel_role

        string Must be one of: master, slave, any

        Sentinel role to use for Redis connections when the redis strategy is defined. Defining this value implies using Redis Sentinel.

      • sentinel_nodes

        array of type record len_min: 1

        Sentinel node addresses to use for Redis connections when the redis strategy is defined. Defining this field implies using a Redis Sentinel. The minimum length of the array is 1 element.

        • host

          string required default: 127.0.0.1

          A string representing a host name, such as example.com.

        • port

          integer default: 6379 between: 0 65535

          An integer representing a port number between 0 and 65535, inclusive.

      • cluster_nodes

        array of type record len_min: 1

        Cluster addresses to use for Redis connections when the redis strategy is defined. Defining this field implies using a Redis Cluster. The minimum length of the array is 1 element.

        • ip

          string required default: 127.0.0.1

          A string representing a host name, such as example.com.

        • port

          integer default: 6379 between: 0 65535

          An integer representing a port number between 0 and 65535, inclusive.

      • ssl

        boolean default: false

        If set to true, uses SSL to connect to Redis.

      • ssl_verify

        boolean default: false

        If set to true, verifies the validity of the server SSL certificate. If setting this parameter, also configure lua_ssl_trusted_certificate in kong.conf to specify the CA (or server) certificate used by your Redis server. You may also need to configure lua_ssl_verify_depth accordingly.

      • server_name

        string

        A string representing an SNI (server name indication) value for TLS.

      • cluster_max_redirections

        integer default: 5

        Maximum retry attempts for redirection.

      • connection_is_proxied

        boolean default: false

        If the connection to Redis is proxied (e.g. Envoy), set it true. Set the host and port to point to the proxy address.

  • response_streaming

    string default: allow Must be one of: allow, deny, always

    Whether to ‘optionally allow’, ‘deny’, or ‘always’ (force) the streaming of answers via server sent events.

  • max_request_body_size

    integer default: 8192

    max allowed body size allowed to be introspected

  • model_name_header

    boolean default: true

    Display the model name selected in the X-Kong-LLM-Model response header

  • targets

    array of type record required
    • route_type

      string required Must be one of: llm/v1/chat, llm/v1/completions, preserve

      The model’s operation implementation, for this provider. Set to preserve to pass through without transformation.

    • auth

      record
      • header_name

        string referenceable

        If AI model requires authentication via Authorization or API key header, specify its name here.

      • header_value

        string referenceable encrypted

        Specify the full auth header value for ‘header_name’, for example ‘Bearer key’ or just ‘key’.

      • param_name

        string referenceable

        If AI model requires authentication via query parameter, specify its name here.

      • param_value

        string referenceable encrypted

        Specify the full parameter value for ‘param_name’.

      • param_location

        string Must be one of: query, body

        Specify whether the ‘param_name’ and ‘param_value’ options go in a query string, or the POST form/JSON body.

      • azure_use_managed_identity

        boolean default: false

        Set true to use the Azure Cloud Managed Identity (or user-assigned identity) to authenticate with Azure-provider models.

      • azure_client_id

        string referenceable

        If azure_use_managed_identity is set to true, and you need to use a different user-assigned identity for this LLM instance, set the client ID.

      • azure_client_secret

        string referenceable encrypted

        If azure_use_managed_identity is set to true, and you need to use a different user-assigned identity for this LLM instance, set the client secret.

      • azure_tenant_id

        string referenceable

        If azure_use_managed_identity is set to true, and you need to use a different user-assigned identity for this LLM instance, set the tenant ID.

      • gcp_use_service_account

        boolean default: false

        Use service account auth for GCP-based providers and models.

      • gcp_service_account_json

        string referenceable encrypted

        Set this field to the full JSON of the GCP service account to authenticate, if required. If null (and gcp_use_service_account is true), Kong will attempt to read from environment variable GCP_SERVICE_ACCOUNT.

      • aws_access_key_id

        string referenceable encrypted

        Set this if you are using an AWS provider (Bedrock) and you are authenticating using static IAM User credentials. Setting this will override the AWS_ACCESS_KEY_ID environment variable for this plugin instance.

      • aws_secret_access_key

        string referenceable encrypted

        Set this if you are using an AWS provider (Bedrock) and you are authenticating using static IAM User credentials. Setting this will override the AWS_SECRET_ACCESS_KEY environment variable for this plugin instance.

      • allow_override

        boolean default: false

        If enabled, the authorization header or parameter can be overridden in the request by the value configured in the plugin.

    • model

      record required
      • provider

        string required Must be one of: openai, azure, anthropic, cohere, mistral, llama2, gemini, bedrock, huggingface

        AI provider request format - Kong translates requests to and from the specified backend compatible formats.

      • name

        string

        Model name to execute.

      • options

        record

        Key/value settings for the model

        • max_tokens

          integer default: 256

          Defines the max_tokens, if using chat or completion models.

        • input_cost

          number

          Defines the cost per 1M tokens in your prompt.

        • output_cost

          number

          Defines the cost per 1M tokens in the output of the AI.

        • temperature

          number between: 0 5

          Defines the matching temperature, if using chat or completion models.

        • top_p

          number between: 0 1

          Defines the top-p probability mass, if supported.

        • top_k

          integer between: 0 500

          Defines the top-k most likely tokens, if supported.

        • anthropic_version

          string

          Defines the schema/API version, if using Anthropic provider.

        • azure_instance

          string

          Instance name for Azure OpenAI hosted models.

        • azure_api_version

          string default: 2023-05-15

          ‘api-version’ for Azure OpenAI instances.

        • azure_deployment_id

          string

          Deployment ID for Azure OpenAI instances.

        • llama2_format

          string Must be one of: raw, openai, ollama

          If using llama2 provider, select the upstream message format.

        • mistral_format

          string Must be one of: openai, ollama

          If using mistral provider, select the upstream message format.

        • upstream_url

          string

          Manually specify or override the full URL to the AI operation endpoints, when calling (self-)hosted models, or for running via a private endpoint.

        • upstream_path

          string

          Manually specify or override the AI operation path, used when e.g. using the ‘preserve’ route_type.

        • gemini

          record
          • api_endpoint

            string

            If running Gemini on Vertex, specify the regional API endpoint (hostname only).

          • project_id

            string

            If running Gemini on Vertex, specify the project ID.

          • location_id

            string

            If running Gemini on Vertex, specify the location ID.

        • bedrock

          record
          • aws_region

            string

            If using AWS providers (Bedrock) you can override the AWS_REGION environment variable by setting this option.

        • huggingface

          record
          • use_cache

            boolean

            Use the cache layer on the inference API

          • wait_for_model

            boolean

            Wait for the model if it is not ready

    • weight

      integer default: 100 between: 1 65535

      The weight this target gets within the upstream loadbalancer (1-65535).

    • description

      string

      The semantic description of the target, required if using semantic load balancing.

    • logging

      record required
      • log_statistics

        boolean required default: false

        If enabled and supported by the driver, will add model usage and token metrics into the Kong log plugin(s) output.

      • log_payloads

        boolean required default: false

        If enabled, will log the request and response body into the Kong log plugin(s) output.