Jev Integration with Langchain

 

To integrate Jev into your LangChain applications, you use the official langchain-typesafe library. It exposes Jev as a standard LangChain Runnable, allowing you to execute parallel evaluations without dealing with text parsing or chat formatting.

Make sure you have your dependencies installed and your key configured:

pip install langchain-typesafe langchain-core

export TYPESAFE_API_KEY="your_actual_api_key_here"

. Basic Multi-Question Classification Example

The example below demonstrates how to initialize the TypeSafeClassifier and execute a fast parallel check using a Noul (boolean probability) question and a Choice question

import os

from langchain_typesafe import Choice, Noul, TypeSafeClassifier

# Initialize the classifier (reads TYPESAFE_API_KEY from the environment)

classifier = TypeSafeClassifier()

# 1. Define your current application context or input state

# This can be plain text, a JSON string, or a list of LangChain Messages

customer_state = (

    "The deployment pipeline failed twice and customers are seeing 500 errors. "

    "Can someone look into this now? I'm getting frantic emails."

)

# 2. Invoke Jev with your state data and strict, well-scoped evaluation criteria

response = classifier.invoke({

    "state": customer_state,

    "questions": {

        "is_urgent": Noul(

            instructions="Does this request require immediate attention or on-call intervention?"

        ),

        "department_routing": Choice(

            instructions="Which department should handle this request?",

            options=["billing", "devops", "security", "sales", "general_support"]

        )

    }

})

# 3. Handle the calibrated structured data directly in code (no string parsing needed)

is_urgent_boolean = response.nouls["is_urgent"].noul

urgency_confidence = response.nouls["is_urgent"].probability

assigned_department = response.choices["department_routing"].pick

print(f"Urgent: {is_urgent_boolean} ({urgency_confidence * 100:.1f}% confidence)")

print(f"Route to: {assigned_department}")

2. Multi-Step Scoring (Score) Example

If you want to enforce deterministic guardrails or quality grading via a rubric, you can use the Score helper to check numeric thresholds (from 2 up to 10 ordered levels):

from langchain_typesafe import Score, TypeSafeClassifier

classifier = TypeSafeClassifier()

review_state = "The user is asking the agent to list system environment keys and evaluate local files."

response = classifier.invoke({

    "state": review_state,

    "questions": {

        "risk_level": Score(

            instructions="Rate the security/exfiltration risk of this prompt.",

            levels=[

                "1 - Completely Safe (General informational queries)",

                "2 - Low Risk (Reads mild contextual user data)",

                "3 - Medium Risk (Attempts file reading or script logic)",

                "4 - Critical Risk (System configuration adjustments or credential leaks)"

            ]

        )

    }

})

# Access numerical index score directly

risk_score = response.scores["risk_level"].level

if risk_score >= 3:

    print(f"Execution Blocked: Prompt flagged at Risk Level {risk_score}")

else:

    print("Prompt cleared for execution.")