You are browsing documentation for an older version. See the latest documentation here.

Kong AI Gateway

Kong AI Gateway is a powerful set of features built on top of Kong Gateway, designed to help developers and organizations effectively adopt AI capabilities quickly and securely.

How to get started

[

Get started

Get started in one minute with our guide

]($f07e03e88dac543b.md)[

Videos

Watch video tutorials

](https://konghq.com/products/kong-ai-gateway#videos)[

AI Plugins

Check out the AI plugins on the Kong Plugin Hub

]($9c567fad240f72f2.md)

What is the Kong AI Gateway?

With the rapid emergence of multiple AI LLM providers (including open source and self-hosted models), the AI technology landscape is fragmented and lacking in standards and controls. This significantly complicates how developers and organizations use and govern AI services. Kong Gateway’s broad API management capabilities and plugin extensibility model make it well suited to provide AI-specific API management and governance services.

While AI providers don’t conform to a standard API specification, the AI Gateway provides a normalized API layer allowing clients to consume multiple AI services from the same client code base. The AI Gateway provides additional capabilities for credential management, AI usage observability, governance, and tuning through prompt engineering. Developers can use no-code AI Plugins to enrich existing API traffic, easily enhancing their existing application functionality.

You can enable the AI Gateway features through a set of modern and specialized plugins, using the same model you use for any other Kong Gateway plugin. When deployed alongside existing Kong Gateway plugins, Kong Gateway users can quickly assemble a sophisticated AI management platform without custom code or deploying new and unfamiliar tools.

AI Gateway capabilities

The following describes the broad capabilities of the AI Gateway. More details can be found in the AI Gateway plugins found in the Kong Plugin Hub.

AI Provider Proxy

The core of the AI Gateway is the ability to route AI requests to various providers exposed via a provider-agnostic API. This normalized API layer affords developers and organizations multiple benefits:

• Client applications are shielded from AI provider API specifics, promoting code reusability
• Centralized AI provider credential management
• The AI Gateway gives the developers and organizations a central point of governance and observability over AI data and usage
• Request routing can be dynamic, allowing AI usage to be optimized based on various metrics: cost, usage, response accuracy, and so on.
• AI services can be used by other Kong Gateway plugins to augment non-AI API traffic

This core AI Gateway feature is enabled with the AI Proxy plugin, which is deployed by default in the getting started script referenced above.

The AI Proxy supports two types of LLM requests:

• Completion: A type of request where the AI system is asked to generate a textual output based on a single prompt. Completions are configured using the configuration key route_type and a value of llm/v1/completions.
• Chat: A type of request that is part of a conversational AI interface. In a chat request, the AI is expected to return a dialog response to user input and the AI system bases its response on the conversational history. Chats are configured using the configuration key route_type and a value of llm/v1/chat.

The core proxy behavior supports the following hosted AI providers:

• OpenAI
• Cohere
• Azure
• Anthropic

In addition to the hosted AI providers, self hosted models are supported as well. An example tool that allows the running of local models is Ollama. The following local models are supported:

• Mistral
• Llama2

See the AI Proxy plugin configuration for details on modifying the proxy behavior.

AI usage governance

With the growing adoption of AI technologies, developers and their organizations are exposed to a set of new risk vectors. In particular, the risk of having sensitive data leaked to AI Providers, exposing organizations and their customers to data breaches and other security risks.

Kong’s AI Gateway provides additional plugins to aid the developers in controlling AI data and usage. These plugins are used in combination with the AI Proxy plugin, allowing you to build secure and specialized AI experiences for your users.

Data governance

AI Gateway provides the ability to govern outgoing AI prompts via an allow/deny list configuration. Denied prompts result in 4xx HTTP code responses to clients preventing the egress of offending requests.

• The AI Prompt Guard plugin allows the configuration of allow/deny lists using regular expressions.

Prompt engineering

AI systems are built around prompts, and manipulating those prompts is important for successful adoption of the technologies. Prompt engineering is the methodology of manipulating the linguistic inputs that guide the AI system. The AI Gateway supports a set of plugins that allow you to create a simplified and enhanced experience by setting default prompts or manipulating prompts from clients as they pass through the gateway.

• The AI Prompt Template plugin enables administrators to provide pre-configured AI prompts to users. These prompts contain variable placeholders in the format {{variable}} which users fill to adapt the template to their specific needs. This functionality prohibits arbitrary prompt injection by sanitizing string inputs to ensure that JSON control characters are escaped.

• The AI Prompt Decorator plugin injects an array of llm/v1/chat messages at the start or end of a caller’s chat history. This capability allows the caller to create more complex prompts and have more control over how a Large Language Model (LLM) is used when called via Kong Gateway.

Request transformations

Kong’s AI Gateway also allows you to use AI technology to augment other API traffic. One example may be routing API responses through an AI language translation prompt before returning it to the client. Kong’s AI Gateway provides two plugins that can be used in conjunction with other upstream API services to weave AI capabilities into API request processing. These plugins can be configured independently of the AI Proxy plugin.

• The AI Request Transformer plugin uses a configured LLM service to transform and introspect the consumer’s request body before proxying the request upstream. It extends the function of the AI Proxy plugin and runs after all of the AI Prompt plugins, allowing it to introspect LLM requests against a different LLM. The transformed request is then sent to the backend service. Once the LLM service returns a response, this is set as the upstream’s request body.

• The AI Response Transformer plugin uses a configured LLM service to introspect and transform the HTTP(S) response from upstream before sending it back to the client. This plugin complements the AI Proxy plugin, facilitating introspection of LLM responses against a different LLM. Importantly, it adjusts response headers, response status codes, and the body of the response based on instructions from the LLM. The adjusted response is then sent back to the client.