Skip to content

Prompt Declaration Language

LLMs will continue to change the way we build software systems. They are not only useful as coding assistants, providing snippets of code, explanations, and code transformations, but they can also help replace components that could only previously be achieved with rule-based systems. Whether LLMs are used as coding assistants or software components, reliability remains an important concern. LLMs have a textual interface and the structure of useful prompts is not captured formally. Programming frameworks do not enforce or validate such structures since they are not specified in a machine-consumable way. The purpose of the Prompt Declaration Language (PDL) is to allow developers to specify the structure of prompts and to enforce it, while providing a unified programming framework for composing LLMs with rule-based systems.

PDL is based on the premise that interactions between users, LLMs and rule-based systems form a document. Consider for example the interactions between a user and a chatbot. At each interaction, the exchanges form a document that gets longer and longer. Similarly, chaining models together or using tools for specific tasks result in outputs that together form a document. PDL allows users to specify the shape of data in such documents in a declarative way (in YAML), and is agnostic of any programming language. Because of its document-oriented nature, it can be used to easily express a variety of data generation tasks (inference, data synthesis, data generation for model training, etc...).

PDL provides the following features: - Ability to use any LLM locally or remotely via LiteLLM, including IBM's watsonx - Ability to templatize not only prompts for one LLM call, but also composition of LLMs with tools (code and APIs). Templates can encompass tasks of larger granularity than a single LLM call - Control structures: variable definitions and use, conditionals, loops, functions - Ability to read from files and stdin, including JSON data - Ability to call out to code and call REST APIs (Python) - Type checking input and output of model calls - Python SDK - Support for chat APIs and chat templates - Live Document visualization

The PDL interpreter takes a PDL program as input and renders it into a document by execution its instructions (calling out to models, code, etc...).

See below for a quick reference, followed by installation notes and an overview of the language. A more detailed description of the language features can be found in this tutorial.

Quick Reference

PDL Quick Reference

(See also PDF version.)

Pro Tip: When writing loops and conditionals with repeat, for, and if-then-else, start the body of the loop or conditional (then/else) with text in order to see the results of every block in the body. See for example this file.

Interpreter Installation

The interpreter has been tested with Python versions 3.11, 3.12, and 3.13, on macOS and Linux. For Windows, please use WSL.

To install the requirements for pdl, execute the command:

pip install prompt-declaration-language

To install the dependencies for development of PDL and execute all the example, execute the command: 

pip install 'prompt-declaration-language[examples]'

Most examples in this repository use IBM Granite models on Replicate. In order to run these examples, you need to create a free account on Replicate, get an API key and store it in the environment variable: - REPLICATE_API_TOKEN

In order to use foundation models hosted on watsonx via LiteLLM, you need a watsonx account (a free plan is available) and set up the following environment variables: - WATSONX_URL, the API url (set to https://{region}.ml.cloud.ibm.com) of your watsonx instance. The region can be found by clicking in the upper right corner of the watsonx dashboard (for example a valid region is us-south ot eu-gb). - WATSONX_APIKEY, the API key (see information on key creation) - WATSONX_PROJECT_ID, the project hosting the resources (see information about project creation and finding project ID).

For more information, see documentation.

To run the interpreter:

pdl <path/to/example.pdl>

The folder examples contains many examples of PDL programs. They cover a variety of prompting patterns such as CoT, RAG, ReAct, and tool use.

We highly recommend to edit PDL programs using an editor that support YAML with JSON Schema validation. For example, you can use VSCode with the YAML extension and configure it to use the PDL schema. This enables the editor to display error messages when the yaml deviates from the PDL syntax and grammar. It also provides code completion. The PDL repository has been configured so that every *.pdl file is associated with the PDL grammar JSONSchema (see settings). You can set up your own VSCode PDL projects similarly using the following .vscode/settings.json file:

{
    "yaml.schemas": {
        "https://ibm.github.io/prompt-declaration-language/dist/pdl-schema.json": "*.pdl"
    },
    "files.associations": {
        "*.pdl": "yaml",
    }
}

The interpreter executes Python code specified in PDL code blocks. To sandbox the interpreter for safe execution, you can use the --sandbox flag which runs the interpreter in a Docker-compatible container. Without this flag, the interpreter and all code is executed locally. To use the --sandbox flag, you need to have a Docker daemon running, such as Rancher Desktop.

The interpreter prints out a log by default in the file log.txt. This log contains the details of inputs and outputs to every block in the program. It is useful to examine this file when the program is behaving differently than expected. The log displays the exact prompts submitted to models by LiteLLM (after applying chat templates), which can be useful for debugging.

To change the log filename, you can pass it to the interpreter as follows:

pdl --log <my-logfile> <my-example>

We can also pass initial data to the interpreter to populate variables used in a PDL program, as follows:

pdl --data <JSON-or-YAML-data> <my-example>

This can also be done by passing a JSON or YAML file:

pdl --data-file <JSON-or-YAML-file> <my-example>

The interpreter can also output a trace file that is used by the Live Document visualization tool (see Live Document):

pdl --trace <file.json> <my-example>

For more information: 

pdl --help

Overview

In PDL, we can write some YAML to create a prompt and call an LLM:

description: Hello world
text:
- "Hello\n"
- model: replicate/ibm-granite/granite-20b-code-instruct-8k
  parameters:
    stop_sequences: '!'
    temperature: 0

The description field is a description for the program. Field text contains a list of either strings or blocks which together form the text to be produced. In this example, the text starts with the string "Hello\n" followed by a block that calls out to a model. In this case, it is model with id replicate/ibm-granite/granite-20b-code-instruct-8k on Replicate, via LiteLLM, with the indicated parameters: the stop sequence is !, and temperature set to 0. Stop sequences are provided with a comma separated list of strings. The input to the model call is everything that has been produced so far in the program (here "Hello\n").

When we execute this program using the PDL interpreter:

pdl examples/hello/hello.pdl

we obtain the following:

Hello
Hello

where the second Hello was produced by Granite. In general, PDL provides blocks for calling models, Python code, and makes it easy to compose them together with control structures (sequencing, conditions, loops).

A similar example on watsonx would look as follows:

description: Hello world
text:
- Hello,
- model: watsonx/ibm/granite-34b-code-instruct
  parameters:
    decoding_method: greedy
    stop:
    - '!'

Notice the syntactic differences. Model ids on watsonx start with watsonx. The decoding_method can be set to greedy, rather than setting the temperature to 0. Also, stop_sequences are indicated with the keyword stop instead as a list of strings.

A PDL program computes 2 data structures. The first is a JSON corresponding to the result of the overall program, obtained by aggregating the results of each block. This is what is printed by default when we run the interpreter. The second is a conversational background context, which is a list of role/content pairs, where we implicitly keep track of roles and content for the purpose of communicating with models that support chat APIs. The contents in the latter correspond to the results of each block. The conversational background context is what is used to make calls to LLMs via LiteLLM.

The PDL interpreter can also stream the background conversation instead of the result:

pdl --stream background examples/hello/hello.pdl

See the tutorial for more information about the conversational background context and how to use roles and chat templates.

Consider now an example from AI for code, where we want to build a prompt template for code explanation. We have a JSON file as input containing the source code and some information regarding the repository where it came from.

For example, given the data in this JSON file: 

source_code: 
  |
  @SuppressWarnings("unchecked")
  public static Map<String, String> deserializeOffsetMap(String lastSourceOffset) throws IOException {
    Map<String, String> offsetMap;
    if (lastSourceOffset == null || lastSourceOffset.isEmpty()) {    
      offsetMap = new HashMap<>();  
    } else {
      offsetMap = JSON_MAPPER.readValue(lastSourceOffset, Map.class);  
    }
    return offsetMap;
  }
repo_info: 
  repo: streamsets/datacollector
  path: stagesupport/src/main/java/com/.../OffsetUtil.java
  function_name: OffsetUtil.deserializeOffsetMap

we would like to express the following prompt and submit it to an LLM:

Here is some info about the location of the function in the repo.
repo: 
streamsets/datacollector
path: stagesupport/src/main/java/com/.../OffsetUtil.java
Function_name: OffsetUtil.deserializeOffsetMap


Explain the following code:

@SuppressWarnings("unchecked")
public static Map<String, String> deserializeOffsetMap(String lastSourceOffset) throws IOException {
  Map<String, String> offsetMap;
  if (lastSourceOffset == null || lastSourceOffset.isEmpty()) {
    offsetMap = new HashMap<>();
  } else {
    offsetMap = JSON_MAPPER.readValue(lastSourceOffset, Map.class);
  }
  return offsetMap;
}

In PDL, this would be expressed as follows (see file):

description: Code explanation example
defs:
  CODE:
    read: ./data.yaml
    parser: yaml
text:
- "\n${ CODE.source_code }\n"
- model: replicate/ibm-granite/granite-3.0-8b-instruct
  input:
      - |
        Here is some info about the location of the function in the repo.
        repo: 
        ${ CODE.repo_info.repo }
        path: ${ CODE.repo_info.path }
        Function_name: ${ CODE.repo_info.function_name }


        Explain the following code:
        ```
        ${ CODE.source_code }```

In this program we first define some variables using the defs construct. Here CODE is defined to be a new variable, holding the result of the read block that follows. A read block can be used to read from a file or stdin. In this case, we read the content of the file ./data.yaml, parse it as YAML using the parser construct, then assign the result to variable CODE.

Next we define a text, where the first block is simply a string and writes out the source code. This is done by accessing the variable CODE. The syntax ${ var } means accessing the value of a variable in the scope. Since CODE contains YAML data, we can also access fields such as CODE.source_code.

The second block calls a granite model on watsonx via LiteLLM. Here we explicitly provide an input field which means that we do not pass the entire text produced so far to the model, but only what is specified in this field. In this case, we specify our template by using the variable CODE as shown above.

When we execute this program with the PDL interpreter, we obtain the following text:

@SuppressWarnings("unchecked")
public static Map<String, String> deserializeOffsetMap(String lastSourceOffset) throws IOException {
  Map<String, String> offsetMap;
  if (lastSourceOffset == null || lastSourceOffset.isEmpty()) {    
    offsetMap = new HashMap<>();  
  } else {
    offsetMap = JSON_MAPPER.readValue(lastSourceOffset, Map.class);  
  }
  return offsetMap;
}

This Java function, `deserializeOffsetMap`, is designed to deserialize a JSON string into a `Map<String, String>`. Here's a breakdown of what it does:

1. It takes a single argument, `lastSourceOffset`, which is expected to be a JSON string.

2. It initializes a `Map<String, String>` called `offsetMap`.

3. If `lastSourceOffset` is either `null` or an empty string, it creates a new `HashMap` and assigns it to `offsetMap`.

4. If `lastSourceOffset` is not `null` or an empty string, it uses the `JSON_MAPPER` object (which is presumably an instance of a JSON deserialization library like Jackson) to deserialize the JSON string into a `Map<String, String>` and assigns it to `offsetMap`.

5. Finally, it returns the `offsetMap`.

The `@SuppressWarnings("unchecked")` annotation is used to suppress a compile-time warning about the raw use of the `Map` type. This is because the `JSON_MAPPER.readValue` method returns a `Map` object, but the compiler doesn't know that this `Map` will be a `Map<String, String>`. The `unchecked` warning is suppressed to avoid this compile-time warning.

Notice that in PDL variables are used to templatize any entity in the document, not just textual prompts to LLMs. We can add a block to this document to evaluate the quality of the output using a similarity metric with respect to our ground truth. See file:

description: Code explanation example
defs:
  CODE:
    read: ./data.yaml
    parser: yaml
  TRUTH:
    read: ./ground_truth.txt
text:
- "\n${ CODE.source_code }\n"
- model: replicate/ibm-granite/granite-3.0-8b-instruct
  def: EXPLANATION
  input: |
      Here is some info about the location of the function in the repo.
      repo: 
      ${ CODE.repo_info.repo }
      path: ${ CODE.repo_info.path }
      Function_name: ${ CODE.repo_info.function_name }


      Explain the following code:
      ```
      ${ CODE.source_code }```
- |


  EVALUATION:
  The similarity (Levenshtein) between this answer and the ground truth is:
- def: EVAL
  lang: python
  code: |
    import textdistance
    expl = """
    ${ EXPLANATION }
    """
    truth = """
    ${ TRUTH }
    """
    result = textdistance.levenshtein.normalized_similarity(expl, truth)

This program has an input block that reads the ground truth from filename examples/code/ground_truth.txt and assigns its contents to variable TRUTH. It also assigns the output of the model to the variable EXPLANATION, using a def construct. In PDL, any block can have a def to capture the result of that block in a variable. The last block is a call to Python code, which is included after the code field. Notice how code is included here simply as data. We collate fragments of Python with outputs obtained from previous blocks. This is one of the powerful features of PDL: the ability to specify the execution of code that is not known ahead of time. We can use LLMs to generate code that is later executed in the same programming model. This is made possible because PDL treats code as data, like any another part of the document.

When we execute this new program, we obtain the following:

@SuppressWarnings("unchecked")
public static Map<String, String> deserializeOffsetMap(String lastSourceOffset) throws IOException {
  Map<String, String> offsetMap;
  if (lastSourceOffset == null || lastSourceOffset.isEmpty()) {    
    offsetMap = new HashMap<>();  
  } else {
    offsetMap = JSON_MAPPER.readValue(lastSourceOffset, Map.class);  
  }
  return offsetMap;
}

This Java method, `deserializeOffsetMap`, is designed to convert a JSON string into a `Map<String, String>`. Here's a breakdown of the code:

1. The method takes a single argument, `lastSourceOffset`, which is expected to be a JSON string.

2. It initializes a `Map<String, String>` called `offsetMap`.

3. If `lastSourceOffset` is either `null` or an empty string, it creates a new `HashMap` and assigns it to `offsetMap`.

4. If `lastSourceOffset` is not `null` or empty, it uses the `JSON_MAPPER` object (which is presumably an instance of `ObjectMapper` from the Jackson library) to convert the JSON string into a `Map<String, String>` and assigns it to `offsetMap`.

5. Finally, it returns the `offsetMap`.

The `@SuppressWarnings("unchecked")` annotation is used to suppress a potential unchecked warning that might occur if the JSON string does not match the expected `Map<String, String>` type.

EVALUATION:
The similarity (Levenshtein) between this answer and the ground truth is:
0.30199115044247793

PDL allows rapid prototyping of prompts by allowing the user to change prompts and see their immediate effects on metrics. Try it!

Finally, we can output JSON data as a result of this program, as follows:

description: Code explanation example
defs:
  CODE:
    read: ./data.yaml
    parser: yaml
  TRUTH:
    read: ./ground_truth.txt
text:
- model: replicate/ibm-granite/granite-3.0-8b-instruct
  def: EXPLANATION
  contribute: []
  input:
     |
      Here is some info about the location of the function in the repo.
      repo: 
      ${ CODE.repo_info.repo }
      path: ${ CODE.repo_info.path }
      Function_name: ${ CODE.repo_info.function_name }


      Explain the following code:
      ```
      ${ CODE.source_code }```
- def: EVAL
  contribute: []
  lang: python
  code:
    |
    import textdistance
    expl = """
    ${ EXPLANATION }
    """
    truth = """
    ${ TRUTH }
    """
    result = textdistance.levenshtein.normalized_similarity(expl, truth)
- data: 
    input: ${ CODE }
    output: ${ EXPLANATION }
    metric: ${ EVAL }

The data block takes various variables and combines their values into a JSON object with fields input, output, and metric. We mute the output of all the other blocks with contribute set to []. The contribute construct can be used to specify how the result of a block is contributed to the overall result, and the background context. Setting it to result contributes the result of the block to the overall result, but not to the background context. Similarly, setting it to context contributes the result of the block to the background context but not the overall result. By default, the result of every block is contributed to both. For the blocks in the program above, we use a def construct to save the intermediate result of each block.

The output of this program is the corresponding serialized JSON object, with the appropriate treatment of quotation marks. Such PDL programs can be bootstrapped in a bash or Python script or piped into a JSONL file to create data en masse.

PDL Language Tutorial

See PDL Language Tutorial

Live Document Visualizer

PDL has a Live Document visualizer to help in program understanding given an execution trace. To produce an execution trace consumable by the Live Document, you can run the interpreter with the --trace argument:

pdl --trace <file.json> <my-example>

This produces an additional file named my-example_trace.json that can be uploaded to the Live Document visualizer tool. Clicking on different parts of the Live Document will show the PDL code that produced that part in the right pane.

This is similar to a spreadsheet for tabular data, where data is in the forefront and the user can inspect the formula that generates the data in each cell. In the Live Document, cells are not uniform but can take arbitrary extents. Clicking on them similarly reveals the part of the code that produced them.

Best Practices

  1. Template Organization:
  2. Keep templates modular and reusable
  3. Use variables for dynamic content

  4. Document template purpose and requirements

  5. Error Handling:

  6. Validate model inputs/outputs
  7. Include fallback logic

  8. Log intermediate results

  9. Performance:

  10. Cache frequent LLM calls
  11. Use appropriate temperature settings

  12. Implement retry logic for API calls

  13. Security:

  14. Enabling sandbox mode for untrusted code
  15. Validate all inputs
  16. Follow API key best practices

Additional Notes

When using Granite models, we use the following defaults for model parameters (except granite-20b-code-instruct-r1.1): - decoding_method: greedy, (temperature: 0) - max_new_tokens: 1024 - min_new_tokens: 1 - repetition_penalty: 1.05

Also if the decoding_method is sample, then the following defaults are used: - temperature: 0.7 - top_p: 0.85 - top_k: 50

For a complete list of issues see here.

Contributing to the Project

See Contributing to PDL.