Pulsar: Structured LLM Output Parser

Pulsar is the today and next-generation tool to build AI applicatios by providing structured output parsing. It allow use with any model hosted in an API compatible with OpenAIAPI, even if that doesn't support json, or function calling (OpenAI o1, Qwen QwQ, DeepSeek R1 ..) and models multi-modal.

Installation

pip install pulsar-struct

Quick Start

from pulsar.client import Client
from pydantic import BaseModel

# Define your output structure
class UserInfo(BaseModel):
    name: str
    age: int
    interests: List[str]

# Initialize client
client = Client()

# Get structured response
response, metadata = client.chat_completion(
    messages=[{"role": "user", "content": "Extract: John, 25 years old, likes reading and hiking"}],
    model="openai/gpt-4o-mini",
    response_type=UserInfo
)

print(response)  # UserInfo(name='John', age=25, interests=['reading', 'hiking'])

Provider Configuration

Current supported providers:

• OpenAI
• Anthropic
• Gemini
• Groq
• OpenRouter
• Ollama
• OpenAIAPILike

Each provider can be configured with specific options:

# Anthropic
client = AnthropicClient(api_key="your-anthropic-key")

# OpenAI
client = OpenAIClient(api_key="your-openai-key")

# Groq
client = GroqClient(api_key="your-groq-key")

# Ollama
client = OllamaClient() # or set the enpoint using base_url="http://host:port" 
                        # default is "http://localhost:11434"

Advanced Usage

Structured output parsing handle all in background for you:

• Convert LLM responses to Pydantic models
• Automatic type coercion
• Handle imperfect JSON responses
• Support for complex data types (enums, literals, nested models)

Custom Response Types

from typing import List, Optional
from pydantic import BaseModel
from enum import Enum

class Sentiment(Enum):
    POSITIVE = "positive"
    NEGATIVE = "negative"
    NEUTRAL = "neutral"

class Analysis(BaseModel):
    sentiment: Sentiment
    confidence: float
    keywords: List[str]
    summary: Optional[str]

response, metadata = client.chat_completion(
    system="Analyse the products",
    messages=[{"role": "user", "content": "Analyze: Great product, highly recommend!"}],
    model="claude-2",
    response_type=Analysis
)

Endpoint Retries

response, _ = client.chat_completion(
    messages=[...],
    model="gpt-4",
    max_retries=3
)

Streaming Support

client = Client()
for chunk, _ in client.chat_completion(
    messages=[...],
    model="groq/llama-3.3-70b-versatile",
    stream=True
):
    print(chunk, end="")

JSON Parsing Features

• Handle unquoted strings
• Fix missing commas and brackets
• Support for partial objects (streaming)
• Automatic type coercion
• Handle embedded type information