Using the AI Response Transformer Plugin

Overview

The AI Response Transformer plugin is designed to operate in two ways:

As a transformer/security arbiter for your existing upstream APIs
As an extension of another AI Proxy LLM route, inspecting and transforming the responses before sending to the upstream LLM service

The plugin configuration consists of two distinct sections:

The LLM configuration, which uses the same configuration options as the AI Proxy plugin.
The prompt (and additional options) containing the instructions for the LLM, which will transform your request.

See the same LLM block in the context of the AI Proxy plugin and the AI Response Transformer plugin:

AI Proxy

AI Response Transformer

config:
  route_type: "llm/v1/chat"
  auth:
    header_name: "Authorization"
    header_value: "Bearer <azure_key>"  # add your own Azure API key
  model:
      provider: "azure"
      name: "gpt-3.5"
      azure_instance: "azure-openai-service"
      azure_deployment_id: "gpt-3-5-deployment"

llm:
  config:
    route_type: "llm/v1/chat"
    auth:
      header_name: "Authorization"
      header_value: "Bearer <azure_key>"  # add your own Azure API key
    model:
      provider: "azure"
      name: "gpt-3.5"
      azure_instance: "azure-openai-service"
      azure_deployment_id: "gpt-3-5-deployment"

prompt: "Mask all credit card numbers in my JSON message with '*'. Return me ONLY the resulting JSON."

When the plugin is accessed in any scope (global, service, route, or consumer), it always sets the upstream’s response body as the user prompt in a chat message, and then sends it to the configured llm: configuration block for inspection or transformation.

Examples

Transforming existing API traffic

This example uses ai-response-transformer on an existing API, for example, something that you have already developed and maintain internally.

Design the prompt.

For this example, we want to intercept responses from the customers API. On each client response, the plugin needs to forward it to the configured large language model, and ask the LLM to mask all credit card numbers with asterisk characters.

The plugin would be configured like this:
```
 config:
   prompt: >
     Mask all credit card numbers in my JSON message with '*'. Return me ONLY the resulting JSON.
   llm:
     # see `ai-proxy` plugin documentation for compatible fields for the "llm" block
```
Attach the plugin.

Attach the ai-response-transformer plugin to the global level, route, service, or consumer on which you want to inspect/transform all responses.

It can even be used on APIs that already have the ai-request-transformer plugin, or it can be used on its own.

What happens next?

First, an upstream API responds to a client request. For example:

 {
   "user": {
     "name": "Kong User",
     "city": "London",
     "credit_card_no": "1234-5678-9012-3456"
   }
 }

Next, Kong parses this into an llm/v1/chat type message, based on your config.prompt:

 {
   "messages": [
     {
       "role": "system",
       "content": "Mask all credit card numbers in my JSON message with '*'. Return me ONLY the resulting JSON."
     },
     {
       "role": "user",
       "content": "{\n\"user\":{\n\"name\":\"Kong User\",\n\"city\":\"London\"\n\"credit_card_no\":\"1234-5678-9012-3456\"}\n}"
     }
   ]
 }

Finally, it sends this to the configured LLM. On the response, it takes the trailing assistant response back from the LLM, and sets it as the HTTP body that will return to the original client:

 {
   "user": {
     "name": "Kong User",
     "city": "London",
     "credit_card_no": "****-****-****-****"
   }
 }

Extraction patterns

If your LLM is a chatbot type, or is unpredictable in responses, you can configure the additional field transformation_extract_pattern with a (PCRE) regular expression to extract the first match from the LLM’s response.

For example, if you have asked for a JSON response but you know that your LLM may add its own text around your answer, use this extraction pattern to withdraw only the JSON object from the LLM’s response:

config:
  prompt: >
    Mask all credit card numbers in my JSON message with '*'. Return me ONLY the resulting JSON.
  transformation_extract_pattern: '\\{((.|\n)*)\\}'

Setting body, headers, and status code

The AI Response Transformer can modify any of the following response sections independently:

Headers
Status code
Body

This allows the Kong admin to configure the LLM to fully orchestrate the response phase inside Kong Gateway.

Enable this feature by setting the config option parse_llm_response_json_instructions.

Design the prompt.

For this example, we want to intercept responses from our customers API. On each client response, the plugin forwards the response to the configured large language model, and asks the LLM to mask all credit card numbers with asterisk characters.

The plugin would be configured like this:

 config:
   prompt: >
     If my JSON message has the user's name 'Kong User', then return this exact JSON message: 
     {"status": 400, "headers": {"x-failed": "true"}, "body": "VALIDATION_FAILURE"}
   parse_llm_response_json_instructions: true
   llm:
     # see `ai-proxy` plugin documentation for compatible fields for the "llm" block

Attach the plugin.

Attach the ai-response-transformer plugin to the global level, route, service, or consumer on which you want to inspect/transform all responses.

It can even be used on APIs that already have the ai-request-transformer plugin, or it can be used on its own.

What happens next?

First, an upstream API responds to a client request, for example:

 {
   "user": {
     "name": "Kong User",
     "city": "London",
     "credit_card_no": "1234-5678-9012-3456"
   }
 }

Once sent to the LLM, the response will be:

 {
   "status": 400,
   "headers": {
     "x-failed": "true"
   },
   "body": "VALIDATION_FAILURE"
 }

Kong will do the following actions:

Set each response header from the headers object
Set the HTTP status code to the status integer
Set the body to be the body string

前へ AI Response Transformerの基本構成例

次へ AI Response Transformer変更履歴