...

Prompting is the process of sending an advice or question to the AI and to get an answer back which can then be further processed:

In the context of AI language models, prompting refers to the way in which you phrase or structure a question, statement, or request to guide the AI in generating a relevant and accurate responseanswer. The prompt is the input provided to the model that defines the task or the type of information you're looking for.Source: ChatGPT

It is very important to write a good prompt in order to get a reliable useful answer back. The process of finding the best prompt for your use case in order to get the required answer is called Prompt Engineering.

PIPEFORCE provides many tools to write, engineer and execute prompts, pipeline prompts chain them in pipelines and supervise them.

...

In this chapter you will learn about the basic prompt principles and how to executed them.

Simple prompt

One of most generic and simplest use cases is to send an advice to the AI and use the response data in your pipeline. For this you can use the ai.promptagent.sendcall command with a simple prompt question.

Here is an example to return some data from the AI using a pipeline:

pipeline:
  - ai.promptagent.sendcall:      |  
      prompt: "Return the names of the 10 biggest cities in the world as JSON array."

Or as an RESTlike API call:

POST https://host/api/v3/command:ai.agent.call
Return the names of the 10 biggest cities in the world as JSON array.

This will result in an entry like this in the body:

[
    "Tokyo",
    "Delhi",
    "Shanghai",
    "Sao Paulo",
    "Mumbai",
    "Beijing",
    "Mexico City",
    "Osaka",
    "Cairo",
    "Dhaka"
]

...

Prompt with context

You can also apply the prompt with context data. This context data can be set as input to the command:

pipeline:
  - ai.promptagent.sendcall:
      input: |
        [
          "Tokyo",
          "Delhi",
          "Shanghai",
          "Sao Paulo",
          "Mumbai",
          "Beijing",
          "Mexico City",
          "Osaka",
          "Cairo",
          "Dhaka"
        ]
      prompt: |
        Order the given list alphabetically.

...

The input of the command will become the context data. It can be plain text, a file or an URI. In case it is a file (for example a PDF or Word document) or any other supported format, it PIPEFORCE will be automatically converted convert it into an AI compatible format.

...

pipeline:
  - ai.promptagent.sendcall:
      input: $uri:drive:invoice-3662.pdf
      prompt: |
        Check the invoice to ensure it is correct both in terms 
        of content and calculations. If everything is fine, return "OK". 
        If not, provide the reason for the error in one sentence.

Convert data with a prompt

See another You can also convert from one data structure into another using a prompt.

See this example which converts a given XML input to JSON:

pipeline:
  - ai.promptagent.sendcall:
      input: |
        <person>
          <firstName>Max</firstName>
          <lastName>Smith</lastName>
          <age>36</age>
        </person>
      prompt: "Convert to JSON"

...

{
    "person": {
        "firstName": "Max",
        "lastName": "Smith",
        "age": 36
    }
}

...

Filter data using a prompt

You can also use a prompt as a data filter.

Here is an example which uses a data privacy filter:

pipeline:
  - ai.promptagent.sendcall:
      input: |
        {
          "person": {
            "firstName": "Max",
            "lastName": "Smith",
            "age": 36
          }
        }
      prompt: |
          Remove all personal data because of privacy and 
          replace by randomized names and add prefix p_

...

{
    "person": {
        "firstName": "p_Alex",
        "lastName": "p_Johnson",
        "age": 48
    }
}

Prompt variables

You can make prompts more dynamically by using prompt variables.

Inside a prompt you can specify a {{variable}}. This variable will be replaced by its value before it gets sent to the AI. Here is an example:

pipeline:
  - ai.agent.call:
      prompt: "Translate this text to {{language}}: {{text}}

...

"
      variables:
        language: "German"
        text: "Hello world!"

Verifying / testing a prompt result

Sometimes it is required that the prompt result must match a given expectation. For example to ensure it can be further processed without errors or for testing purposes.

For this, you can specify an optional verify message. If given, an additional pompt will be executed which will verify the last prompt result using the given verify message as the verification condition. If verification has been failed, an error will the thrown and pipeline execution will stop.

Here is an example:

pipeline:
  - ai.prompt.send:
      message: "Convert these values into a JSON: firstName=Max, lastName=Smith"
      verify: "The answer is a JSON and contains the fields firstName and lastName."

This approach is also very handy for testing purposes:

pipeline:
  - ai.prompt.send:
      message: "Calculate 10 + 5 and return only the result as number. "
      verify: "The answer is a number and greater than 10."

Require a structured output format

When it comes to the answer to a prompt, sometimes it might be needed to conform to a specific structure like boolean, number or JSON for example, so it can be further processed automatically.

In PIPEFORCE you can define this structure using the answerFormat parameter as this example shows:

pipeline:
  - ai.prompt.send:
      message: "Calculate 10 + 5."
      response:
        type: integer
        schema: 

If you specify the answerFormat, it is not required to add the format advice to the prompt message. This is done automatically for you to ensure only the result in given answer format is returned without any additional text. Furthermore: If the backend LLM in use supports specific output formatting as additional meta data, this will be automatically applied for you to as it is even more reliable to enure a specific structure is returned.

Send multiple messages

In case you need to send multiple messages in one prompt, you can use the parameter messages like this:

...

If both parameters, messages and prompt message are given, prompt message will be automatically added to messages of role system user at the very end.

Possible values for role are:

...

The parameter content can be a plain text or any AI convertable object (like a PDF file for example). The conversion and preparation to an AI compatible format is done by PIPEFORCE automatically.

Extract data from text

A special form of a prompt is an extractor. It will extract a specific data format from a given prompt and makes sure the result complies exactly with this data format so it can be directly further processed.

Extract date and time from text

This example extracts a date and time from a given text:

pipeline:
  - ai.extract.datetime:
      message: |
        It happened in the evening of 1968, just fifteen minutes 
        shy of midnight, following the celebrations of Independence Day.
      pattern: yyyy-MM-dd, hh:mm:ss

Result:

1968-07-04, 23:45:00

The parameter pattern is optional. If missing the date and time will be formatted using UTC format: YYYY-MM-DDTHH:MM:SSZ

Since the current date and time in UTC is passed automatically with the prompt, you can also ask relative time questions like this example shows:

pipeline:
  - ai.extract.datetime:
      message: |
        I'm 24 years old and my birthday is on April 19th. When was my day of birth?
      pattern: yyyy-MM-dd

Result:

2000-04-19

Extract sentiment (mood)

In this example you can detect wether a given message has a negative, neutral or positive mood.

pipeline:
  - ai.extract.sentiment:
      message: I love AI.

Result:

positive

Extract boolean

This example checks whether a given text is a true or false thing:

pipeline:
  - ai.extract.boolean:
      message: All animals have feet.

Result:

false

Extract JSON

This example extracts a JSON from a given text:

pipeline:
  - ai.extract.json:
      message: |
        My name is Max Smith. I'm 39 years old and I live in Los Angeles, CA.

Result:

{
  "name": "Max Smith",
  "age": 39,
  "location": {
    "city": "Los Angeles",
    "state": "CA"
  }
}

Extract geopraphical location

This example extracts a geographical location from a given text and returns it a JSON with country, city, latitude and longitude information:

pipeline:
  - ai.extract.location:
      message: |
        I strolled through the streets of the city, past the radiant orange trees 
        and the futuristic buildings of the Ciudad de las Artes y las Ciencias.

Result:

{
  "country": "Spain",
  "city": "Valencia",
  "latitude": "39.4550° N",
  "longitude": "0.3546° W"
}

Extract people

It is also possible to extract all people mentioned in a given text:

pipeline:
  - ai.extract.people:
      message: |
        Mike had a conversation with Sarah about the upcoming music fest.

Result:

[
  {"name": "Mike"},
  {"name": "Sarah"}
]