Skip to content

Latest commit

 

History

History
198 lines (145 loc) · 6.32 KB

base-connectors.md

File metadata and controls

198 lines (145 loc) · 6.32 KB
Raw

Connectors

🔍 Extract Connector

Extracts information from structured or unstructured data in multiple input variables, potentially doing simple conversions along the way, and stores the result in one or more output variables.

Configuration

Provide a JSON object schema of the structure to extract from the input, with descriptions of what they should contain:

{
  product: {
      description: "name of the product",
      type: "string"
  },
  price: {
      description: "price of the product",
      type: "number"
  },
  tags: {
      description: "tags for the product",
      type: "array",
      items: {
        type: "string",
        description: "a product tag",
        enum: ["A", "B", "C"]
      }
    }
}

The following types are supported:

  • string
  • integer
  • number
  • boolean
  • object
  • array

Note

The non-LLM models currently do not support the array type.

For details, see here.

You should always provide a speaking name and description and be aware that both are essentially part of the prompt engineering and determine how well the information is extracted.

As seen above, you can provide an enum array if all possible values are known.

Since extracting string values is very common, there is a shorthand:

{
  firstname: "first name",
  lastname: "last name"
}

This is equivalent to:

{
  firstname: {
      description: "first name",
      type: "string"
  },
  lastname: {
      description: "last name",
      type: "string"
  }
}

Extract a list of entities

Select Extraction Mode Multiple Entities to extract a list of multiple entity objects, each conforming to the configured schema. You can provide an optional description for the entities to extract.

Note

For non-LLM models, extracting multiple entities is experimental and requires the entity description.

The result will be a list of objects or an empty list.

Result

A temporary variable result that contains a result JSON object of the same form as configured above. Can be mapped to one or more process variables using the result expression.

⚖ Decide Connector

Makes decisions based on multiple input variables and stores the result decision (and potentially the reasoning behind it) in output variables.

Configuration

Provide a natural language question or description of what the connector should decide, e.g.:

What is the intention of the customer's mail?

Next, select the Output Type (Boolean, Integer, Float or String). If not Boolean, you may restrict the connector to a classification on a finite set of options, instead of letting it freely choose the result value:

[
  "CANCEL_SUBSCRIPTION",
  "CHANGE_SUBSCRIPTION",
  "COMPLAINT",
  "OTHER"
]

Result

A temporary variable result that contains a result JSON object with a field decision containing the final decision and - for LLMs - a field reasoning containing an explanation of the reasoning behind the decision. Can be mapped to one or more process variables using the result expression.

Note

The non-LLM models do not provide a reasoning.

For details, see here.

✍🏼 Compose Connector

Composes texts for emails, letters, chat messages, or social media posts based on multiple input variables and stores the result text in an output variable.

Configuration

General Properties

Configure a desired text type, style, tone, language, and length for the text.

Variance

Select a Variance value (controls model temperature). None will make the output mostly deterministic and more focused. The higher the variance, the more diverse and unpredictable the text becomes. A higher value is a good fit for creating creative content that should change on every run. Select None if you want to be as precise as possible and don't need diverse outputs.

Template

The template dictates the shape of the result text. You can use template variables using curly braces. Variables that are present in the input variable mapping are replaced directly, without going through the model. The remaining template variables will be filled in by the LLM. You can use simple variable names if obvious enough, or write full sentences with instructions on what to fill in. The result text is the template with all variables replaced or filled in.

Example:

Hello {name},

{ thank the customer for his purchase }

Yours,
{agentName}

Here name and agentName could be input variables, while the middle part would be generated. name and agentName will not be sent to the model in any way (be careful that the template variable names correctly match the input variable names). Whitespaces at the beginning and end of a variable are ignored and therefore optional.

Result

A temporary variable result that directly contains the result text. Can be mapped to a process variable using the result expression.

🌍 Translate Connector

Translates multiple input variables to any given language and stores the result in one or more output variables

Configuration

Enter the target language (e.g. English).

Note

The non-LLM translation model currently supports the following languages: DANISH, DUTCH, ENGLISH, FINNISH, FRENCH, GERMAN, ITALIAN, NORWEGIAN, POLISH, PORTUGUESE, SPANISH, SWEDISH, UKRAINIAN

For details, see here.

Result

A temporary variable result that contains a result JSON object with a field for every input field, containing the translation. Can be mapped to one or more process variables using the result expression.

🪄 Generic Connector

Can execute custom tasks not covered by the specialized connectors.

Configuration

Describe the task:

Perform task X and store the result in the result field. Also describe the reasoning behind your result.

Specify the output schema:

{
  result: "the result of the task",
  reasoning: "the reasoning behind the task result"
}

Result

A temporary variable result that contains a result JSON object as specified in the output schema. Can be mapped to one or more process variables using the result expression.