Structured, Ecto outputs with OpenAI (and OSS LLMs)

Structured prompting for LLMs. Instructor is a spiritual port of the great Instructor Python Library by @jxnlco, check out his talk on YouTube.

The Instructor library is useful for coaxing an LLM to return JSON that maps to an Ecto schema that you provide, rather than the default unstructured text output. If you define your own validation logic, Instructor can automatically retry prompts when validation fails (returning natural language error messages to the LLM, to guide it when making corrections).

Instructor is designed to be used with the OpenAI API by default, but it also works with llama.cpp and Bumblebee (Coming Soon!) by using an extendable adapter behavior.

At its simplest, usage is pretty straightforward:

1. Create an ecto schema, with a @doc string that explains the schema definition to the LLM.
2. Define a validate_changeset/1 function on the schema, and use the Instructor.Validator macro in order for Instructor to know about it.
3. Make a call to Instructor.chat_completion/1 with an instruction for the LLM to execute.

You can use the max_retries parameter to automatically, iteratively go back and forth with the LLM to try fixing validation errorswhen they occur.

defmodule SpamPrediction do
  use Ecto.Schema
  use Instructor.Validator

  @doc """
  ## Field Descriptions:
  - class: Whether or not the email is spam.
  - reason: A short, less than 10 word rationalization for the classification.
  - score: A confidence score between 0.0 and 1.0 for the classification.
  """
  @primary_key false
  embedded_schema do
    field(:class, Ecto.Enum, values: [:spam, :not_spam])
    field(:reason, :string)
    field(:score, :float)
  end

  @impl true
  def validate_changeset(changeset) do
    changeset
    |> Ecto.Changeset.validate_number(:score,
      greater_than_or_equal_to: 0.0,
      less_than_or_equal_to: 1.0
    )
  end
end

is_spam? = fn text ->
  Instructor.chat_completion(
    model: "gpt-3.5-turbo",
    response_model: SpamPrediction,
    max_retries: 3,
    messages: [
      %{
        role: "user",
        content: """
        Your purpose is to classify customer support emails as either spam or not.
        This is for a clothing retail business.
        They sell all types of clothing.

        Classify the following email: 
        ```
        #{text}
        ```
        """
      }
    ]
  )
end

is_spam?.("Hello I am a Nigerian prince and I would like to send you money")

# => {:ok, %SpamPrediction{class: :spam, reason: "Nigerian prince email scam", score: 0.98}}

Check out our Quickstart Guide for more code snippets that you can run locally (in Livebook). Or, to get a better idea of the thinking behind Instructor, read more about our Philosophy & Motivations.

Optionally, you can also customize the your llama.cpp calls (with defaults shown):

llamacpp
config :instructor, adapter: Instructor.Adapters.Llamacpp
config :instructor, :llamacpp,
    chat_template: :mistral_instruct,
    api_url: "http://localhost:8080/completion"

In your mix.exs,

def deps do
  [
    {:instructor, "~> 0.0.5"}
  ]
end

• llamacpp adapter broken, needs to support openai input/output API
  • GBNF should enforce required properties on objects, currently they're optional.
  • GBNF limit the number of digits in number tokens -- small models can sometimes run off to infinit digits
• Add instructor tests against llamacpp interface using mocks, there's non-trivial logic in there
• Logging for Distillation / Finetuning
• Add a Bumblebee adapter
• Add llamacpp_ex adapter
• Support naked ecto types by auto-wrapping, not just maps of ecto types, do not wrap if we don't need to... Current codepaths are muddled
• Support Streaming
  • Verify schemaless support {:array, %{name: :string}}
  • Support typespec style support for array streaming [MySchema]
• Optional/Maybe types
• Add Livebook Tutorials, include in Hexdocs
  • Text Classification
  • Self Critique
  • Image Extracting Tables
  • Moderation
  • Citations
  • Knowledge Graph
  • Entity Resolution
  • Search Queries
  • Query Decomposition
  • Recursive Schemas
  • Table Extraction
  • Action Item and Dependency Mapping
  • Multi-File Code Generation
  • PII Data Sanitizatiommersed
• Update hexdocs homepage to include example for tutorial
• Setup Github CI for testing, add badge to README

• Why structured prompting?

  Meditations on new HCI. Finally we have software that can understand text. f(text) -> text. This is great, as it gives us a new domain, but the range is still text. While we can use string interpolation to map Software 1.0 into f(text), the outputs are not interoperable with Software 1.0. Hence why UXs available to us are things like Chatbots as our users have to interpret the output.

  Instructor, structure prompting, gives use f(text) -> ecto_schema. Schemas are the lingua franca of Software 1.0. With Instrutor we can now seamlessly move back and forth between Software 1.0 and Software 2.0.

  Now we can maximally leverage AI...

• From GPT-4 to zero-cost production - Distilation, local-llms, and the cost structure of AI.

  ... 😘