LangChainKit

LangChainKit simplifies the process of prompting LLMs to return structured outputs using LangChain and LangFuse.

🚀 Features

• 🔧 Simplified Qwen3 + vLLM integration
  Automatically configure enable_thinking and other complex settings for Qwen3 models when using vLLM.

• 🧠 Structured Output via LangChain
  Easily prompt the LLM to generate structured outputs, including batch prompting support, with minimal setup.

• 📊 LangFuse Integration
  Track and evaluate LLM performance using LangFuse, without writing boilerplate code.

Installation

pip install langchainkit

Quick Start

Configuration

Set up your environment variables in .env file:

DEEPSEEK_API_KEY=your deepseek api key
MOONSHOT_API_KEY=...
OPENROUTER_API_KEY=...
ARK_API_KEY=...
DASHSCOPE_API_KEY=...
LOCAL_VLLM_BASE_URL=http://172.20.14.28:8000/v1
LOCAL_VLLM_API_KEY=...

LANGFUSE_SECRET_KEY=...
LANGFUSE_PUBLIC_KEY=...
LANGFUSE_HOST=...

from langchainkit import GeneralLLM,prompt_parsing
from pydantic import BaseModel
from dotenv import load_dotenv

load_dotenv() # load .env file

llm = GeneralLLM.deepseek_chat()

class Response(BaseModel):
    answer: str
    confidence: float

result = prompt_parsing(
    model=Response,
    failed_model=Response(answer="no_answer", confidence=0.0),
    query="What is the capital of France?",
    llm=llm,
    use_langfuse=False 
)
print(result.answer)  # "Paris"
print(result.confidence)  # 1.0

result = prompt_parsing(
    model=Response,
    failed_model=Response(answer="no_answer", confidence=0.0),
    query=["What is the capital of France?",
           "What is the capital of Germany?",
           "What is the capital of Italy?"],
    llm=llm,
    use_langfuse=False
)
for each in result:
    print(each.answer)
    print(each.confidence)
# Paris
# 0.95
# Berlin
# 0.95
# Rome
# 1.0

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• LangChain for the core framework
• vLLM for high-throughput LLM inference
• Langfuse for observability and monitoring