logillm 0.2.0

A generic, high-performance, low-dependency LLM programming framework inspired by dspy

LogiLLM

Program language models with structured inputs and outputs, built for production.

LogiLLM lets you define what you want from an LLM using Python signatures, then handles prompt engineering, output parsing, and optimization automatically.

Quick Example

from logillm.core.predict import Predict
from logillm.providers import create_provider, register_provider

# Setup (one time)
provider = create_provider("openai")
register_provider(provider, set_default=True)

# Define what you want
analyzer = Predict("text -> sentiment: str")

# Use it
result = await analyzer(text="I love this new feature!")
print(result.sentiment)  # "positive"

Instead of crafting prompts and parsing responses, you specify the input/output structure. LogiLLM handles the rest.

What You Can Build

📧 Smart Email Processing - Classify, extract entities, determine urgency
📝 Document Generation - Create structured reports from data
🤖 Reasoning Agents - Multi-step analysis with external tool integration
🎮 Interactive Systems - Games, chatbots, personalized experiences
💾 Persistent AI - Systems that remember and learn from interactions

Key Features

• 🎯 Structured I/O - Type-safe inputs and outputs with automatic validation
• ⚡ Zero Dependencies - Core library uses only Python standard library
• 🚀 Production Ready - Full async support, comprehensive error handling, observability
• 🔧 Auto-Optimization - Improve performance by optimizing prompts AND hyperparameters
• 💾 Built-in Persistence - Save and load optimized models instantly
• 📊 JSONL Logging - Track optimization runs with complete reproducibility

Installation

# Core library 
pip install logillm

# With LLM providers
pip install logillm[openai]     # For GPT models
pip install logillm[anthropic]  # For Claude models

Learning Path

📚 Complete Documentation

• Getting Started Guide - Personalized tutorial recommendations
• Tutorial Matrix - Difficulty levels and time estimates
• Tutorial Examples - Complete working code for all tutorials

→ Full Tutorial Index - 6 complete tutorials from beginner to advanced

Working Examples

Basic Classification

# Classify text into categories
classifier = Predict("text -> category: str")
result = await classifier(text="I need help with my billing account")
print(result.category)  # "Billing"

Multi-Output Prediction

# Get multiple outputs in one call
analyzer = Predict("text -> sentiment: str, confidence: float")
result = await analyzer(text="This product is amazing!")
print(result.sentiment)   # "positive"  
print(result.confidence)  # 0.97

Complex Type Support (New!)

# Work with lists, dictionaries, and optional types
extractor = Predict("document: str -> entities: dict[str, list[str]], summary: Optional[str]")
result = await extractor(document="Apple Inc. announced iPhone 15 in Cupertino.")
print(result.entities)  # {"companies": ["Apple Inc."], "products": ["iPhone 15"], "locations": ["Cupertino"]}

# Use Union types with pipe syntax
processor = Predict("data: str | bytes -> processed: bool, format: str")

Multimodal Capabilities (New!)

from logillm.core.signatures.types import Image, Audio, Tool

# Vision analysis
vision = Predict("image: Image -> description: str, objects: list[str], confidence: float")
result = await vision(image=Image.from_path("photo.jpg"))
print(result.objects)  # ["person", "laptop", "coffee"]

# Audio processing
transcriber = Predict("audio: Audio -> transcript: str, language: str")
result = await transcriber(audio=Audio.from_url("https://example.com/speech.mp3"))

# Tool/Function calling
tool_agent = Predict("query: str -> tool_calls: list[Tool], explanation: str")

Structured Signatures with Field Validation (Enhanced!)

from logillm.core.signatures import Signature, InputField, OutputField

class TextAnalysis(Signature):
    # Required and optional fields with defaults
    text: str = InputField(description="Text to analyze")
    max_length: int = InputField(default=100, description="Maximum response length")
    
    # Output fields with type hints
    category: str = OutputField(description="Category like support, sales, billing")
    priority: int = OutputField(description="Priority 1-5 where 5 is most urgent")
    entities: list[str] = OutputField(description="Named entities found")

analyzer = Predict(signature=TextAnalysis)
result = await analyzer(text="My account is locked and I cannot access my files")
# Automatically validates inputs and outputs against field specifications
print(result.category)  # "Account Access"
print(result.priority)  # 1

Production Features

Multiple Providers

# Set up multiple LLM providers  
openai_provider = create_provider("openai", model="gpt-4.1")
register_provider(openai_provider, "fast")

# Use the registered provider (becomes default)
result = await Predict("text -> summary: str")(text="Long document to summarize...")
print(result.summary)  # "This document discusses..."

Optimization (requires training data)

# Optimize performance on your data
from logillm.optimizers import HybridOptimizer
from logillm.core.optimizers import AccuracyMetric

# Your labeled training examples
training_data = [
    {"inputs": {"text": "Great service!"}, "outputs": {"sentiment": "positive"}},
    {"inputs": {"text": "Poor quality"}, "outputs": {"sentiment": "negative"}},
    # ... more examples
]

# Optimize both prompts and hyperparameters  
classifier = Predict("text -> sentiment: str")
optimizer = HybridOptimizer(metric=AccuracyMetric(key="sentiment"))
result = await optimizer.optimize(module=classifier, dataset=training_data)

# Save optimized model
result.optimized_module.save("sentiment_model.json")

# Load in production
classifier = Predict.load("sentiment_model.json")

Optimizer Comparison

LogiLLM provides a comprehensive suite of optimizers, each designed for different optimization scenarios:

Optimizer	Type	What it Optimizes	Best For	Key Features
HybridOptimizer ⭐	Hybrid	Prompts + Hyperparameters	Production systems	LogiLLM's killer feature - simultaneous optimization
SIMBA	Hybrid	Hyperparameters + Demos	Complex tasks	Introspective rule generation with mini-batches
MIPROv2	Prompt	Instructions + Demos	Multi-stage optimization	Bayesian optimization with multi-objective support
COPRO	Prompt	Instructions	Instruction refinement	Breadth-first search with temperature control
BootstrapFewShot	Prompt	Few-shot examples	Learning from data	Teacher-student demonstration generation
KNNFewShot	Prompt	Example selection	Dynamic examples	Semantic similarity-based selection
FormatOptimizer	Prompt	Output format	Format discovery	Tests JSON vs XML vs Markdown
InstructionOptimizer	Prompt	Task instructions	Instruction clarity	LLM-based instruction generation
LabeledFewShot	Prompt	Hand-crafted examples	Baseline comparison	Traditional few-shot approach
HyperparameterOptimizer	Hyperparameter	Temperature, top_p, etc.	Parameter tuning	Bayesian, Grid, Random search
AvatarOptimizer	Ensemble	Multiple personas	Complex reasoning	Multi-perspective ensemble
ReflectiveEvolution	Prompt	Execution traces	Self-improvement	LLM reflection on past runs
MultiObjective	Hybrid	Multiple metrics	Trade-off optimization	Balance accuracy, cost, latency

Optimizer Taxonomy

graph LR
    subgraph "Prompt Optimizers"
        direction TB
        P1[BootstrapFewShot<br/>📚 Teacher-Student Learning]
        P2[COPRO<br/>🔄 Instruction Refinement]
        P3[KNNFewShot<br/>🎯 Semantic Selection]
        P4[FormatOptimizer<br/>📝 JSON/XML/Markdown]
        P5[InstructionOptimizer<br/>💡 LLM Generation]
        P6[LabeledFewShot<br/>✋ Manual Examples]
    end
    
    subgraph "Hyperparameter Optimizers"
        direction TB
        H1[HyperparameterOptimizer<br/>🎛️ Bayesian Search]
        H2[GridSearch<br/>📊 Systematic]
        H3[RandomSearch<br/>🎲 Stochastic]
    end
    
    subgraph "Hybrid Optimizers"
        direction TB
        Y1[HybridOptimizer ⭐<br/>🚀 Prompts + Params]
        Y2[SIMBA<br/>🧠 Rules + Params]
        Y3[MIPROv2<br/>🔀 Multi-stage Pipeline]
        Y4[MultiObjective<br/>⚖️ Balance Metrics]
    end
    
    subgraph "Specialized"
        direction TB
        S1[AvatarOptimizer<br/>👥 Multi-persona]
        S2[ReflectiveEvolution<br/>🔍 Self-improvement]
    end
    
    style Y1 fill:#f96,stroke:#333,stroke-width:3px
    style Y2 fill:#fcc,stroke:#333,stroke-width:2px
    style Y3 fill:#fcc,stroke:#333,stroke-width:2px

Optimization Workflow

graph LR
    A[📊 Training Data] --> B{Choose Optimizer}
    B -->|Simple Task| C[Prompt Only]
    B -->|Parameter Tuning| D[Hyperparameter Only]
    B -->|Best Results| E[Hybrid Strategy ⭐]
    
    E --> F{Select Strategy}
    F -->|Alternating| G[Prompts → Params → Repeat]
    F -->|Joint| H[Simultaneous Optimization]
    F -->|Sequential| I[Params → Then Prompts]
    
    C --> J[Evaluate Performance]
    D --> J
    G --> J
    H --> J
    I --> J
    
    J --> K{Good Enough?}
    K -->|No| L[Adjust & Retry]
    K -->|Yes| M[💾 Save Model]
    
    L --> B
    M --> N[🚀 Deploy to Production]
    
    N --> O[Load & Use]
    O --> P[Monitor Performance]
    
    style E fill:#f96,stroke:#333,stroke-width:3px
    style M fill:#9f9,stroke:#333,stroke-width:2px
    style N fill:#9f9,stroke:#333,stroke-width:2px

Choosing the Right Optimizer

For most use cases: Start with HybridOptimizer - it optimizes both prompts and hyperparameters:

optimizer = HybridOptimizer(
    metric=your_metric,
    strategy="alternating"  # or "joint", "sequential"
)

For specific scenarios:

• Limited training data? → Use LabeledFewShot with hand-crafted examples
• Need dynamic examples? → Use KNNFewShot for semantic similarity selection
• Complex multi-step tasks? → Use MIPROv2 for sophisticated pipeline optimization
• Want to understand why it works? → Use SIMBA for introspective rule generation
• Multiple competing objectives? → Use MultiObjective to balance trade-offs

JSONL Logging for Optimization Tracking

# Track and analyze optimization runs
from logillm.core.jsonl_logger import OptimizationLogger

logger = OptimizationLogger(filepath="optimization.jsonl")
result = await logger.log_optimization(
    optimizer=optimizer,
    module=classifier,
    dataset=training_data,
    validation_set=validation_data
)

# Analyze the log
import json
with open("optimization.jsonl") as f:
    events = [json.loads(line) for line in f]
    
# See score progression, hyperparameters, prompts, and more
for event in events:
    if event['event_type'] == 'evaluation_end':
        print(f"Score: {event['score']:.2%}")

→ Full JSONL Logging Documentation

Why LogiLLM?

Coming from prompt engineering? Stop writing brittle string templates. Define structured inputs/outputs and let LogiLLM handle prompt construction.

Coming from dspy? Get better performance through hybrid optimization (prompts + hyperparameters), zero-dependency deployment, and more modern Python standards.

Building production systems? Native async support, comprehensive error handling, automatic retries, and observability built-in.

What's Different from DSPy?

• Hybrid Optimization - Optimize both prompts AND hyperparameters simultaneously
• Zero Dependencies - Core framework uses only Python standard library
• Production First - Built for scaling, monitoring, and maintenance from day one
• Type Safety - Full Pydantic integration with IDE support
• Modern Python - Async/await throughout, Python 3.9+ features

Getting Help

• 📚 Start with Tutorials - 6 hands-on tutorials, 15-40 minutes each
• 🤔 Getting Started Guide - Personalized recommendations
• ⚡ Quick Examples - Working code for common patterns
• 🐛 Issues - Report bugs or request features on GitHub

---

Ready to build? Start with the LLM Text Generation tutorial (15 minutes) or browse all tutorials to find your perfect starting point.