dslmodel 2024.11.19

Pydantic + DSPy instances from prompts and Jinja.

DSLModel

pip install dslmodel

Intro Video: Welcome to DSLModel

Custom GPT: DSLModel Assistant v2024.10.10

Overview

DSLModel is a powerful Python framework for declarative model creation using templates and concurrent execution. Built atop Pydantic for data validation and DSPy for model execution, DSLModel streamlines the development of dynamic models with features like:

• Dynamic Field Generation: Utilize Jinja2 templates for flexible model definitions.
• Concurrent Execution: Leverage concurrent processing to optimize performance.
• Workflow Management: Define and execute complex workflows with conditional logic and loops.
• Finite State Machines: Incorporate state machine patterns for managing state transitions.
• Data Handling Utilities: Read from and write to various data formats seamlessly.
• AI-Assisted Development: Enhance productivity with AI-driven tools for code generation.

Table of Contents

• DSLModel
  • Custom GPT: DSLModel Assistant v2024.10.10
  • Overview
  • Table of Contents
  • Installation
  • Getting Started
    • Defining Models
    • Generating Data from Templates
    • Concurrent Execution
    • Workflow Management
    • Workflow YAML
    • Finite State Machines
    • Data Handling
  • Architecture
    • Core Components
    • Data Flow
  • Development
    • Setup
    • Testing
    • Contributing
  • Deployment
    • Deployment Pipeline
  • License
  • Contact

Installation

Ensure you have Python 3.12 or higher installed. Then, install DSLModel via pip:

pip install dslmodel

Alternatively, install from source:

git clone https://github.com/your-username/dslmodel.git
cd dslmodel
poetry install

Getting Started

Defining Models

Create dynamic models using Jinja2 templates and DSLModel.

from typing import List
from pydantic import Field
from dslmodel import DSLModel


class Participant(DSLModel):
    """Represents a participant in a meeting."""
    name: str = Field("{{ fake_name() }}", description="Name of the participant.")
    role: str = Field("{{ fake_job() }}", description="Role of the participant.")


class Meeting(DSLModel):
    """Represents a meeting and its participants."""
    name: str = Field(..., description="Name of the meeting.")
    participants: List[Participant] = Field(..., description="List of participants.")

Generating Data from Templates

Use templates to generate model instances with dynamic content.

from typing import List, Optional, Dict, Union
from pydantic import Field

from dslmodel import init_lm, DSLModel


class Participant(DSLModel):
    """Represents a participant in a meeting."""
    name: str = Field("{{ fake_name() }}", description="Name of the participant.")
    role: str = Field("{{ fake_job() }}", description="Role of the participant.")


class Meeting(DSLModel):
    """Represents a meeting, its participants, agenda, and other details."""
    name: str = Field(..., description="Name of the meeting.")
    meeting_date: str = Field(..., description="Date of the meeting.")
    location: Optional[str] = Field(None, description="Location where the meeting is held.")
    chairperson: Participant = Field(..., description="Chairperson of the meeting.")
    secretary: Participant = Field(..., description="Secretary responsible for taking minutes.")
    participants: List[Participant] = Field(..., description="List of all participants in the meeting.")
    agenda: List[str] = Field(..., description="Agenda items for the meeting.", min_length=3)
    minutes: List[str] = Field(..., description="Minutes of the meeting. Time, Description", min_length=3)
    rules_of_order: List[str] = Field(..., description="Rules governing the meeting.", min_length=3)


participants = [Participant() for _ in range(5)]  # Created using Jinja defaults (no LM)

# Generate the Meeting 
init_lm()  # Sets the lm to gpt-4o-mini

meeting_template = """
Fortune 500 Meeting about {{ fake_bs() }}
Participants:
{% for participant in participants %}
- {{ participant.name }} ({{ participant.role }})
{% endfor %}
"""

meeting_instance = Meeting.from_prompt(meeting_template, participants=participants)

print(meeting_instance.to_yaml())

Concurrent Execution

Execute multiple tasks concurrently to improve performance.

from dslmodel import init_lm, DSLModel
from dslmodel.utils.model_tools import run_dsls
from pydantic import Field


class Participant(DSLModel):
    """Represents a participant in a meeting."""
    name: str = Field(..., description="Name of the participant.")
    role: str = Field(..., description="Role of the participant.")


tasks = [(Participant, "Create a person with a name and job role") for _ in range(5)]

init_lm()  # Sets the lm to gpt-4o-mini

results = run_dsls(tasks, max_workers=5)

for i, result in enumerate(results):
    print(f"Participant {i + 1}: {result}")

Workflow Management

Define and execute complex workflows using Workflow, Job, and Action.

from dslmodel.workflow import Workflow, Job, Action, Condition, CronSchedule

condition = Condition(expr="len(participants) > 3")

action1 = Action(
    name="Generate Participants",
    code="participants.extend([Participant() for _ in range(5)])"
)

action2 = Action(
    name="Notify Organizer",
    code="print('Organizer notified.')",
    cond=condition
)

job = Job(
    name="Setup Meeting",
    steps=[action1, action2]
)

trigger = CronSchedule(cron="0 9 * * MON")  # Every Monday at 9 AM

workflow = Workflow(
    name="Weekly Meeting Setup",
    triggers=[trigger],
    jobs=[job],
    context={"participants": []}
)

workflow.execute()

Workflow YAML

workflow:
  name: "Weekly Meeting Setup"
  triggers:
    - type: "CronSchedule"
      cron: "0 9 * * MON"  # Every Monday at 9 AM
  context:
    participants: [ ]
  jobs:
    - name: "Setup Meeting"
      steps:
        - name: "Generate Participants"
          code: "participants.extend([Participant() for _ in range(5)])"
        - name: "Notify Organizer"
          code: "print('Organizer notified.')"
          condition:
            expr: "len(participants) > 3"

Finite State Machines

Manage state transitions using FSMMixin.

import logging
from enum import Enum, auto
from dslmodel.mixins import FSMMixin, trigger


class SalesState(Enum):
    INITIALIZING = auto()
    RESEARCHING = auto()
    OUTREACHING = auto()
    CLOSING = auto()
    COMPLETING = auto()


class ChallengerSalesAgent(FSMMixin):
    def __init__(self):
        super().__init__()
        self.setup_fsm(state_enum=SalesState, initial=SalesState.INITIALIZING)

    @trigger(source=SalesState.INITIALIZING, dest=SalesState.RESEARCHING)
    def start_research(self):
        print("Starting market research.")

    @trigger(source=SalesState.RESEARCHING, dest=SalesState.OUTREACHING)
    def conduct_outreach(self):
        print("Conducting outreach to leads.")

    @trigger(source=SalesState.OUTREACHING, dest=SalesState.CLOSING)
    def close_deal(self):
        print("Closing the deal with the client.")

    @trigger(source=SalesState.CLOSING, dest=SalesState.COMPLETING)
    def complete_sale(self):
        print("Completing the sale.")

    def forward(self, prompt, **kwargs):
        super().forward(prompt, **kwargs)
        print(f"Processing prompt: {prompt}")


def main():
    agent = ChallengerSalesAgent()
    print("Initial state:", agent.state)

    # Simulating the simplified flow of the sales process
    agent.forward("start researching the market")
    print("State after research:", agent.state)

    agent.forward("reach out to leads")
    print("State after outreach:", agent.state)

    agent.forward("finalize the deal")
    print("State after closing the deal:", agent.state)

    agent.forward("complete the sale")
    print("Final state:", agent.state)


if __name__ == '__main__':
    main()

Data Handling

Read from and write to various data formats using DataReader and DataWriter.

from dslmodel import DataReader, DataWriter

# Reading data
data_reader = DataReader(file_path="data/sample_data.csv")
data = data_reader.forward()
print(data)

# Writing data
data_writer = DataWriter(data=data, file_path="output/data_output.csv")
data_writer.forward()

Architecture

Core Components

• DSLModel: Core framework for declarative model creation using templates.
• Mixins:
  • ToolMixin: Adds dynamic tool execution capabilities.
  • FSMMixin: Provides finite state machine functionality.
• Workflow Components:
  • Workflow, Job, Action, Condition, CronTrigger: Orchestrate complex workflows.
• Data Handling Utilities:
  • DataReader, DataWriter: Handle data ingestion and output.

Data Flow

User Inputs -> DSLModel Templates -> Generated Models -> Validation and Execution

Development

Setup

1. Clone the Repository

   git clone https://github.com/your-username/dslmodel.git
   cd dslmodel
   

2. Install Dependencies

   poetry install
   

3. Configure Environment Variables

   Create a .env file and add necessary environment variables, such as OPENAI_API_KEY.

4. Run the Development Server

   poe api --dev
   

Testing

Run tests using pytest:

poetry run pytest

Ensure test coverage is at least 90%.

Contributing

Contributions are welcome! Please follow the contribution guidelines and adhere to the code of conduct.

Deployment

DSLModel utilizes GitHub Actions for continuous integration and deployment.

Deployment Pipeline

1. Code Push: Developers push code to the repository.
2. Automated Testing: GitHub Actions run test suites.
3. Linting: Code is linted using ruff to maintain quality.
4. Build and Deployment: Successful builds are deployed to staging or production.

License

Distributed under the MIT License. See LICENSE for more information.

Contact

• Project Link: https://github.com/seanchatmangpt/dslmodel
• Issues: https://github.com/seanchatmangpt/dslmodel/issues

---

By following this guide, you can effectively utilize DSLModel for declarative model creation, workflow management, data handling, state machine implementation, and AI-assisted development.