Best practices: How to contribute to files?

[afuetterer]

Hi there,

I am a new user of this project. I could not find information on how to output LLM responses to a file in the docs.

Then I read in #1341, that it is possible to contribute to files. The example in the release notes is a log file. Should this be used for outputting JSON or other typed / structured output as well? Could you maybe add an example to the tutorials?

[vazirim]

Hi @afuetterer, thanks for your interest in PDL!

Here's an example that takes the output from an LLM, creates JSON data and outputs the structured data to a file:

defs:
  myfile:
    aggregator: 
      file: myfile.json
text:
- Hello
- model: ollama_chat/granite3.3:8b
  def: out
- data:
    response: ${ out }
  contribute:
  - myfile

Hope this helps. Please let us know if you have any further questions.

[afuetterer]

This looks promising, let me give it a try. I think I will need to combine it with a spec and parser:json?

[afuetterer]

Hi @vazirim, I am not sure what I am doing wrong. I added a minimal example with a dummy prompt that generates a list of three objects, in this case dinosaurs for fun:

defs:
  myfile:
    aggregator:
      file: dinosaurs.json
text:
- |
    Give me a list of three dinosaurs.
    Describe them as json onjects, using the keys common_name, scientific_name and diet.
    Only generate the list as json.
- model: ollama/granite4
  def: out
- data: ${ out }
  parser: json
  spec: {type: array, minItems: 1, maxItems: 3, items: { type: object, properties: { common_name: {type: string}, scientific_name: {type: string}, diet: {type: string} } }}
  contribute:
  - myfile

It runs without errors, that is great. But the output in the file dinosaurs.json looks like a list of python dicts.

[{'common_name': 'Tyrannosaurus rex', 'scientific_name': 'Tyrannosaurus rex', 'diet': 'carnivore'}, {'common_name': 'Triceratops', 'scientific_name': 'Triceratops horridus', 'diet': 'herbivore'}, {'common_name': 'Velociraptor', 'scientific_name': 'Velociraptor mongoliensis', 'diet': 'carnivore'}]

The output of the pdl run looks like this:

$ uv run pdl dinosaurs.pdl
Give me a list of three dinosaurs.
Describe them as json onjects, using the keys common_name, scientific_name and diet.
Only generate the list as json.
Here is the list of three dinosaurs in JSON format:

[
  {
    "common_name": "Tyrannosaurus rex",
    "scientific_name": "Tyrannosaurus rex",
    "diet": "carnivore"
  },
  {
    "common_name": "Triceratops",
    "scientific_name": "Triceratops horridus",
    "diet": "herbivore"
  },
  {
    "common_name": "Velociraptor",
    "scientific_name": "Velociraptor mongoliensis",
    "diet": "carnivore"
  }
][{'common_name': 'Tyrannosaurus rex', 'scientific_name': 'Tyrannosaurus rex', 'diet': 'carnivore'}, {'common_name': 'Triceratops', 'scientific_name': 'Triceratops horridus', 'diet': 'herbivore'}, {'common_name': 'Velociraptor', 'scientific_name': 'Velociraptor mongoliensis', 'diet': 'carnivore'}]

Am I creating a python list of dicts with the parsing and spec and outputting that to the file now? VS Code complains, that a JSON file should have double quotes.

I feel like I am half way there. Could you help out?

[vazirim]

@afuetterer Thank you for pointing this out! We will fix this and make a release shortly.

[afuetterer]

Thanks a lot, I tried with the above example (dinosaurs). The output is now valid JSON.

[mandel]

Thank you @afuetterer!

The release 0.9.1 that is now available should fix the issue.