Skip to content

Latest commit

 

History

History
106 lines (65 loc) · 5.85 KB

CORE.md

File metadata and controls

106 lines (65 loc) · 5.85 KB

💙 Core of the Promptbook

🏛 Organization

Organization groups together workflows, workforce, knowledge, instruments, and actions into one package. Entities in one organization can share resources (i.e. import workflows, teams, personas, knowledge, instruments and actions from each other).

Each organization has a unique URL; for example, https://promptbook.studio/my-cool-project/.

🏗 Workflow

A workflow represents a piece of work that has specific input and output.

Private workflows can be imported within the organization, while public workflows can be imported inside the organization or used everywhere through their unique URLs. Each workflow has a unique URL; for example, https://promptbook.studio/my-cool-project/workflows/generate-website.

As a programmer, you can imagine a workflow as an async function that can be used inside your code. It exposes an interface with a record of input and output parameters but hides the internal implementation.

You can use workflows in other workflows, use it in classic programming languages as async functions that can be called, run workflow in CLI, or use Promptbook Studio to create instant miniapps.

Task

A task is one step in a workflow. Each task is divided into two parts: the actual task job and a check that the result of the job is correct.

Each task can use results from previous tasks. Tasks in a workflow form a directed acyclic graph.

Task job

A task job is the actual work that needs to be done. It can be:

  • Asking a persona to do a job; this is a higher abstraction of the calling model
  • Asking the user to do a job
  • Searching the knowledge base for information
  • Using an action to perform external work and get the result
  • Using an instrument, for example, a calculator
  • A simple template to concatenate parameters to a hardcoded template
  • Script execution to run custom code (Python/JavaScript/TypeScript/...)
  • Calling a model directly to perform a job, bypassing the organization's workforce—for example, calling GPT's Assistant

Task check

After the task job is done, the result can be checked to see if it is correct. If it is not correct, the task is repeated a certain number of times.

You can expect:

  • Result is in the expected format; for example, an available domain name
  • Result is in the expected range; for example, between 1 sentence and 2 paragraphs
  • Adversarial check by another persona to approve the result

🏋️‍♂️ Workforce

The workforce is an abstraction above LLM models, tokens, temperature, top-k, top-p, and other model parameters. You can describe what you desire in human language like Jane, creative writer with a sense of sharp humor instead of gpt-4-2024-13-31, temperature 1.2, top-k 40, STOP token ".\n",....

Persona

A persona is the basic unit of the workforce. It is defined by its description; for example, Jane, creative writer with a sense of sharp humor.

This persona description is used to select the best model and parameters for the job. If the persona has conflicting requirements, Promptbook will try to find the best compromise or even combine multiple models to achieve the best result.

For example, Josh, lawyer with perfect language and logic capabilities and a strong sense of privacy is not possible to achieve with one model. Big models like GPT-4 or Claude-3.5 are great for language and logic, but they send data to the cloud. On the other hand, LLAMA-3 is great for privacy but not as strong in language and logic. Therefore, Promptbook will create an ad-hoc meta-model using LLAMA-3 to strip all data, GPT-4 for language and logic, and then back to LLAMA-3 to reintroduce sensitive data.

Each persona can have access to different knowledge, instruments, and actions.

Team

A team groups personas together. A team can also group other teams to form a complex responsibility hierarchy. Each team can have access to different knowledge, instruments, and actions.

Role

A role is an ad-hoc modification of a persona. A role can be defined for a specific task; for example, Jane (email writer). Roles have access to the same knowledge, instruments, and actions as their parent persona.

💡 Knowledge

Knowledge is external information that is used in task jobs. Knowledge can be:

  • Explicit as a text directly in the workflow, team, or persona
  • File which is referenced from the workflow, team, or persona We support various file types like pdf, docx, txt, md, odt, doc, rtf, and it's possible to easily add support for other file types. The file is parsed and stored in the knowledge base
  • Website which is referenced from the workflow, team, or persona The website is scraped and stored in the knowledge base

Knowledge piece

A knowledge piece is the smallest unit of knowledge that makes sense on its own. Every type of knowledge—explicit text, file, or website—is parsed/scraped and divided into knowledge pieces. These pieces are indexed, put into the knowledge base, and can be used in task jobs via techniques like retrieval-augmented generation.

🛠 Instruments

Instruments are external information that cannot be pre-scraped and need to be fetched at the moment of the task job. For example:

  • Current Time and date
  • User's Location
  • Searching the internet
  • Computing some mathematical expression
  • Weather in some location
  • Stock price of some company
  • Availability of some Domain name
  • Calling a GET endpoint of an API

This is an abstraction above function calling and API calling in models.

☎ Actions

Actions are similar to instruments but can change the state of the world. For example:

  • Sending email
  • Creating a file
  • Ending a workflow
  • Calling a POST endpoint of an API