The `ChatClient` offers a fluent API for stateless interaction with an AI Model. It supports both a synchronous and reactive programming model.
The `ChatClient` offers a fluent API for communicating with an AI Model.
It supports both a synchronous and reactive programming model.
The fluent API has methods for building up the constituent parts of a xref:api/prompt.adoc#_prompts[Prompt] that is passed to the AI model as input.
The fluent API has methods for building up the constituent parts of a xref:api/prompt.adoc#_prompt[Prompt] that is passed to the AI model as input.
The `Prompt` contains the instructional text to guide the AI model's output and behavior. From the API point of view, prompts consist of a collection of messages.
The AI model processes two main types of messages: user messages, which are direct inputs from the user, and system messages, which are generated by the system to guide the conversation.
These messages often contain template placeholders that are substituted at runtime based on user input to customize the response of the AI model to the user input.
These messages often contain placeholders that are substituted at runtime based on user input to customize the response of the AI model to the user input.
There are also Prompt options that can be specified., such as the name of the AI Model to generate content and the temperature setting that controls the randomness or creativity of the generated output.
There are also Prompt options that can be specified., such as the name of the AI Model to use and the temperature setting that controls the randomness or creativity of the generated output.
== Creating a ChatClient
@@ -69,8 +70,8 @@ The ChatClient API offers several ways to format the response from the AI Model.
=== Returning a ChatResponse
The response from the AI model is a rich structure defined by the type xref:api/chatmodel.adoc#_chatresponse[ChatResponse].
It includes metadata about how the response was generated and can also contain multiple responses, known as xref:api/chatmodel.adoc#_generation[Generation]s, each with its own metadata.
The response from the AI model is a rich structure defined by the type xref:api/chatmodel.adoc#ChatResponse[ChatResponse].
It includes metadata about how the response was generated and can also contain multiple responses, known as xref:api/chatmodel.adoc#Generation[Generation]s, each with its own metadata.
The metadata includes the number of tokens (each token is approximately 3/4 of a word) used to create the response.
This information is important because hosted AI models charge based on the number of tokens used per request.
@@ -124,6 +125,7 @@ The `stream` lets you get an asynchronous response as shown below
You can also stream the `ChatResponse` using the method `Flux<ChatResponse> chatResponse()`.
TIP: The chunked stream response can not be converted into a Java entities automatically.
Use the xref:api/structured-output-converter.adoc#_structuredoutputconverter[Structured Output Converter] to convert the aggregated response.
In the 1.0.0 M2 we will offer a convenience method that will let you return an Java entity with the reactive `stream()` method.
In the meantime, you should use the xref:api/structured-output-converter.adoc#StructuredOutputConverter[Structured Output Converter] to convert the aggregated response explicity as shown below.
This also demonstrates the use of parameters in the fluent API that will be discussed in more detail in a later section of the documentation.
== Using defaults and parameters
[source,java]
----
var converter = new BeanOutputConverter<>(new ParameterizedTypeReference<List<ActorsFilms>>() {
});
It is often useful to create a `ChatClient` with default user and/or system text defined at design time.
By design time, we mean creating it at application startup in a `@Configuration` class.
You can then replace the user or system parameters in the main runtime code.
Flux<String> flux = this.chatClient.prompt()
.user(u -> u.text("""
Generate the filmography for a random actor.
{format}
""")
.param("format", converter.getFormat()))
.stream()
.content();
Here is a simple example which defines the system text without using any parameters.
The system text sets the context, instructions, and overall desired behavior for the model's response
After specifying the `call` method on `ChatClient` there are a few different options for the response type.
* `String content()`: returns the String content of the response
* `ChatResponse chatResponse()`: returns the `ChatResponse` object that contains multiple generations and also metadata about the response, for example how many token were used to create the response.
* `entity` to return a Java type
** entity(ParameterizedTypeReference<T> type): used to return a Collection of entity types.
** entity(Class<T> type): used to return a specific entity type.
** entity(StructuredOutputConverter<T> structuredOutputConverter): used to specify an instance of a `StructuredOutputConverter` to convert a `String` to an entity type.
You can also invoke the `stream` method instead of `call` and
== stream() return values
After specifying the `stream` method on `ChatClient`, there are a few options for the response type:
* `Flux<String> content()`: Returns a Flux of the string being generated by the AI model.
* `Flux<ChatResponse> chatResponse()`: Returns a Flux of the `ChatResponse` object, which contains additional metadata about the response.
== Using Defaults
Creating a ChatClient with default system text in an `@Configuration` class simplifies runtime code.
By setting defaults, you only need to specify user text when calling `ChatClient, eliminating the need to set system text for each request in your runtime codeala path.
=== Default System Text
In the following example, we will configure the system text to always reply in a pirate's voice.
To avoid repeating the system text in runtime code, we will create a `ChatClient` instance in an `@Configuration` class.
return builder.defaultSystem("You are a friendly chat bot that answers question in the voice of a {voice}")
.build();
}
}
----
[source,java]
----
@RestController
class AIController {
private final ChatClient chatClient
AIController(ChatClient chatClient) {
this.chatClient = chatClient;
}
@GetMapping("/ai")
Map<String, String> completion(@RequestParam(value = "message", defaultValue = "Tell me a joke") String message, String voice) {
return Map.of(
"completion",
chatClient.prompt()
.system(sp -> sp.param("voice", voice))
.user(message)
.call()
.content());
}
}
----
The response is
[source.bash]
----
http localhost:8080/ai voice=='Robert DeNiro'
{
"completion": "You talkin' to me? Okay, here's a joke for ya: Why couldn't the bicycle stand up by itself? Because it was two tired! Classic, right?"
}
----
=== Other defaults
At the `ChatClient.Builder` level, you can specify the default prompt.
* `defaultOptions(ChatOptions chatOptions)`: Pass in either portable options defined in the `ChatOptions` class or model-specific options such as those in `OpenAiChatOptions`. For more information on model-specific `ChatOptions` implementations, refer to the JavaDocs.
* `defaultFunction(String name, String description, java.util.function.Function<I, O> function)`: The `name` is used to refer to the function in user text. The `description` explains the function's purpose and helps the AI model choose the correct function for an accurate response. The `function` argument is a Java function instance that the model will execute when necessary.
* `defaultFunctions(String... functionNames)`: The bean names of `java.util.Function`s defined in the application context.
* `defaultUser(String text)`, `defaultUser(Resource text)`, `defaultUser(Consumer<UserSpec> userSpecConsumer)`: These methods let you define the user text. The `Consumer<UserSpec>` allows you to use a lambda to specify the user text and any default parameters.
* `defaultAdvisors(RequestResponseAdvisor... advisor)`: Advisors allow modification of the data used to create the `Prompt`. The `QuestionAnswerAdvisor` implementation enables the pattern of `Retrieval Augmented Generation` by appending the prompt with context information related to the user text.
* `defaultAdvisors(Consumer<AdvisorSpec> advisorSpecConsumer)`: This method allows you to define a `Consumer` to configure multiple advisors using the `AdvisorSpec`. Advisors can modify the data used to create the final `Prompt`. The `Consumer<AdvisorSpec>` lets you specify a lambda to add advisors, such as `QuestionAnswerAdvisor`, which supports `Retrieval Augmented Generation` by appending the prompt with relevant context information based on the user text.
You can override these defaults at runtime using the corresponding methods without the `default` prefix.
A common pattern when calling an AI model with user text is to append or augment the prompt with contextual data.
This contextual data can be of different types. Common types include:
* **Your own data**: This is data the AI model hasn't been trained on. Even if the model has seen similar data, the appended contextual data takes precedence in generating the response.
* **Conversational history**: The chat model's API is stateless. If you tell the AI model your name, it won't remember it in subsequent interactions. Conversational history must be sent with each request to ensure previous interactions are considered when generating a response.
=== Retrieval Augmented Generation
A vector database stores data that the AI model is unaware of.
When a user question is sent to the AI model, a `QuestionAnswerAdvisor` queries the vector database for documents related to the user question.
The response from the vector database is appended to the user text to provide context for the AI model to generate a response.
Assuming you have already loaded data into a `VectorStore`, you can perform Retrieval Augmented Generation (RAG) by providing an instance of `QuestionAnswerAdvisor` to the `ChatClient`.
Is this example, the `SearchRequest.defaults()` will perform a similarity search over all documents in the Vector Database.
To restrict the types of documents that are searched, the `SearchRequest` takes a SQL like filter expression that is portable across all `VectorStores`.
=== Chat Memory
The interface `ChatMemory` represents a storage for chat conversation history. It provides methods to add messages to a
* conversation, retrieve messages from a conversation, and clear the conversation history.
There is one implementation `InMemoryChatMemory` that provides in-memory storage for chat conversation history.
Two advisor implementations use the `ChatMemory` interface to advice the prompt with conversation history which differ in the details of how the memory is added to the prompt
* `MessageChatMemoryAdvisor` : Memory is retrieved added as a collection of messages to the prompt
* `PromptChatMemoryAdvisor` : Memory is retrieved added into the prompt's system text.
* `VectorStoreChatMemoryAdvisor` : The construtor ` VectorStoreChatMemoryAdvisor(VectorStore vectorStore, String defaultConversationId, int chatHistoryWindowSize)` lets you specify the VectorStore to retrieve the chat history from, the unqiue conversation ID, the size of the chat history to be retreived in token size.
A sample `@Service` implementaiton that uses several advisors is shown below
Testing AI applications requires evaluating the generated content to ensure the AI model has not produced a hallucinated response.
One method to evaluate the response is to use the AI model itself for evaluation. Select the best AI model for the evaluation, which may not be the same model used to generate the response.
The Spring AI interface for evaluating responses is `Evaluator`, defined as:
The input to the evaluation is the `EvaluationRequest` defined as
[source,java]
----
public class EvaluationRequest {
private final String userText;
private final List<Content> dataList;
private final ChatResponse chatResponse;
public EvaluationRequest(String userText, List<Content> dataList, ChatResponse chatResponse) {
this.userText = userText;
this.dataList = dataList;
this.chatResponse = chatResponse;
}
...
}
----
* `userText`: The raw input from the user.
* `dataList`: Contextual data, such as from Retrieval Augmented Generation, appended to the raw input.
* `chatResponse`: The AI model's response.
== RelevancyEvaluator
One implementation is the `RelevancyEvaluator`, which uses the AI model for evaluation. More implementations will be available in future releases.
The `RelevancyEvaluator` uses the input (`userText`) and the AI model's output (`chatResponse`) to ask the question:
[source, text]
----
Your task is to evaluate if the response for the query
is in line with the context information provided.\n
You have two options to answer. Either YES/ NO.\n
Answer - YES, if the response for the query
is in line with context information otherwise NO.\n
Query: \n {query}\n
Response: \n {response}\n
Context: \n {context}\n
Answer: "
----
Here is an example of a JUnit test that performs a RAG query over a PDF document loaded into a Vector Store and then evaluates if the response is relevant to the user text.
[source,java]
----
@Test
void testEvaluation() {
dataController.delete();
dataController.load();
String userText = "What is the purpose of Carina?";
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Mark Pollack