netanel-core 0.2.0

Self-learning LLM call library. Every call learns. Every call improves.

netanel-core

The self-learning LLM call. Every call learns. Every call improves.

A Python library that wraps any LLM with file-based memory and automatic quality improvement. No database required.

from netanel_core import LearningLLM

llm = LearningLLM(namespace="my-app")
result = llm.call("Write a function to validate emails")

print(result.response)  # The LLM's output
print(result.score)      # Quality score (0.0-1.0)
print(result.usage)      # Token usage for cost tracking

---

✨ Features

• 🧠 Self-Learning - Extracts patterns from every call, builds better context over time
• 📊 Quality-First - Auto-evaluation + retry loop, only stores high-quality outputs
• 💾 File-Based Memory - No database, human-readable .md files, git-trackable
• 🔄 Prompt Evolution - Auto-rewrites prompts based on learnings
• 🎯 Bounded Safety - Max retries, tokens, iterations (NASA-grade)
• 🤖 DeepAgent Support - Complex reasoning with LangGraph agents
• 💰 Cost Tracking - Token usage for all calls

---

🚀 Quick Start

pip install netanel-core

from netanel_core import LearningLLM

llm = LearningLLM(namespace="my-app")
result = llm.call("Explain quantum computing simply")

if result.passed:
    print(result.response)
    print(f"Quality: {result.score:.2f}, Tokens: {result.usage['total_tokens']}")

---

📖 How It Works

Every llm.call() executes:

1. RETRIEVE - Load memories from namespace
2. BUILD - Create context (role + memories + task)
3. CALL - Invoke LLM
4. EVALUATE - Score quality (gpt-4o-mini + main model)
5. RETRY - If score < threshold, retry with feedback
6. EXTRACT - Extract patterns from successful responses
7. STORE - Save to memories/{namespace}/patterns/
8. EVOLVE - Trigger prompt improvements
9. RETURN - Result with response + metadata

---

💡 Usage Examples

Custom Model

from langchain_openai import ChatOpenAI

model = ChatOpenAI(model="gpt-4", temperature=0.7)
llm = LearningLLM(namespace="gpt4-app", model=model)

DeepAgent Mode

result = llm.call(
    "Research AI papers and summarize trends",
    use_agent=True  # Multi-step reasoning
)
print(f"Steps: {result.agent_steps}")

Cost Tracking

total = sum(llm.call(task).usage['total_tokens'] for task in tasks)
cost = (total / 1_000_000) * 0.375  # gpt-4o-mini avg
print(f"Cost: ${cost:.4f}")

---

🎯 Configuration

from netanel_core import Config

config = Config(
    namespace="my-app",
    quality_threshold=0.8,
    max_retries=3,
    memories_dir="./memories",
)

llm = LearningLLM(config=config)

Or YAML:

namespace: my-app
quality_threshold: 0.8
max_retries: 3

config = Config.from_yaml("config.yaml")

---

📂 Memory Structure

memories/
└── {namespace}/
    ├── patterns/
    │   ├── 001-function-structure.md
    │   └── 002-error-handling.md
    └── prompts/
        └── current.md

Human-readable Markdown - inspect or version control.

---

🧪 Development

pip install -e .[dev]
pytest --cov

---

📚 Documentation

• Architecture - System design
• API Reference - Complete API
• Examples - Usage patterns

---

📝 License

MIT - see LICENSE

---

Built by Netanel Systems with LangGraph + Deep Agents