AI Form Filler Models
AI Form Filler provides two built-in LLM implementations using ChatGPT services: ChatGPTService and ChatGPTChatCompletionService.
Both models are quite good at understanding and processing natural language. In any case, it is advised to try both implementations to find out which one fits better with your requirements. Even if ChatGPTChatCompletionService, using the "gpt-3.5-turbo-16k-0613" model, is more advanced, the LLM world isn’t 100% predictable. You never know what model could work better.
Implementing Custom AI Model Adapter
You can implement your own AI model adapter for ChatGPT or any other LLM provider. You should use one of the existing implementations as a template if you want to create a new adaptor. Internal parameters such as the model name, temperature, or max tokens are up to you to choose and parametrize. The methods you need to implement are the ones described by the LLMService interface.
First, you need to implement the getPromptTemplate() method to create the prompt that is sent to the AI module. This method receives all the internal structures generated automatically from the target form. Then you need to implement the getGeneratedResponse() method that sends the prompt to the AI module. This method has to guarantee that the response is a valid JSON object using the same format that is sent in the request.
/**
* An LLM service that generates a response based on a prompt.
* All responsibilities related to the model usage have to be
* implemented in this service. This could be APIKEY providing,
* parameter setting, prompt template generation, etc.
*/
public interface LLMService {
/**
* Generates a prompt based on the input, the target components, and any
* extra instruction.
*
* @param input the input text (e.g. "My name is John")
* @param objectMap the objectMap containing the target components in a
* hierarchical structure (keys = ids)
* @param typesMap the map containing the desired type to fill with each
* one of the target components (keys = ids)
* @param componentInstructions the instructions containing additional
* information to understand the purpose of the
* fields.
* @param contextInstructions the context instructions containing additional
* information to understand the input or
* providing some information not present in the input
* @return the generated prompt to be sent to the AI module
*/
public String getPromptTemplate(String input, Map<String, Object> objectMap, Map<String, String> typesMap, HashMap<Component, String> componentInstructions, ArrayList<String> contextInstructions);
/**
* Generates a response based on the input prompt from the AI module.
*
* @param prompt the prompt to be used by the AI module
* @return the generated response from the AI module. This response has to
* be a valid JSON Object using target field IDs as keys and the value
* correctly formatted according to the target component. An example is
* the following format:
* <br>
* <pre>{@code
* {
* "field Id 1": "value 1", // TextField
* "field Id 2": 66, // NumberField or IntegerField
* "field Id 3": "2022-04-03", // DatePicker
* "field Id 4 ": [ // Grid
* {
* "inner item id1": "Value 1",
* "inner item id2": "Value 2",
* "inner item id3": "Value 3"
* },
* {
* "inner item id1": "Value 1",
* "inner item id2": "Value 2",
* "inner item id3": "Value 3"
* }
* ],
* "field Id 5 ": ["Value 1", "Value 2"], // MultiSelectComboBox
* "field Id 6": true, // Checkbox
* "field Id 7": 43.47 // BigDecimalField
* }
* }</pre> <br>
*/
public String getGeneratedResponse(String prompt);
}Then pass your newly implemented service to the Form Filler constructor.
FormFiller formFiller = new FormFiller(formLayout, new MyGPTService());
FormFillerResult result = formFiller.fill(input);