# TokenCut + LangChain

One callback handler to compress all prompts in your LangChain pipeline.

## Installation

```
pip install agentready-sdk langchain-openai
```

## Basic Usage

```
from langchain_openai import ChatOpenAI
from agentready.integrations.langchain import TokenCutCallbackHandler

# Create the handler (use your API key from env)
import os
handler = TokenCutCallbackHandler(api_key=os.environ["AGENTREADY_API_KEY"])

# Attach to any LangChain LLM
llm = ChatOpenAI(model="gpt-4", callbacks=[handler])

# All prompts are automatically compressed
response = llm.invoke("Your very long context here...")

# Check cumulative savings
print(handler.stats)
# {'total_tokens_saved': 3421, 'total_savings_usd': 0.1026}
```

## With RAG Chain

```
from langchain.chains import RetrievalQA

handler = TokenCutCallbackHandler(api_key=os.environ["AGENTREADY_API_KEY"], level="aggressive")

qa_chain = RetrievalQA.from_chain_type(
    llm=ChatOpenAI(model="gpt-4", callbacks=[handler]),
    retriever=vectorstore.as_retriever(),
)

# RAG context is compressed before hitting GPT-4
answer = qa_chain.invoke("What is the system architecture?")
print(f"Saved {handler.stats['total_tokens_saved']} tokens")
```

## Configuration

```
handler = TokenCutCallbackHandler(
    api_key=os.environ["AGENTREADY_API_KEY"],
    level="medium",         # light, medium, aggressive
    preserve_code=True,     # keep code blocks intact
    min_length=400,         # skip short prompts (chars)
    compress_system=False,  # skip system messages
)
```

## Why Use This?

- **Zero code changes:** Just add a callback — no need to modify your chains
- **RAG optimization:** Compress retrieved context before LLM processing
- **Cost tracking:**`handler.stats` gives cumulative savings
- **Safe:** Falls back to original text if compression fails