conversational-graph-dev 0.0.14

Conversational graph workflow package

Conversational Graph

A Python package for building stateful, graph-based conversational agents. Control the flow of your AI conversations with precision using states, transitions, and tools.

Installation

pip install conversational-graph-dev

Features

• Graph-Based Flow Control: Define conversations as a graph of states and transitions.
• State Management: Manage conversation context and strict progression between logical steps.
• Tool Integration: Integrate HTTP request tools to specific states.
• Data Modeling: Strongly typed data extraction using Pydantic models.

Quick Start

Here is a simple example of how to build a strict conversational flow for a loan assistant.

from conversational_graph import ConversationalGraph, ConversationalDataModel, PreDefinedTool, HttpToolRequest

get_weather = PreDefinedTool().http_tool(HttpToolRequest(
        name="get_weather",
        description="Called when the user asks about the weather. This function will return the weather for the given location.",
        url="https://api.open-meteo.com/v1/forecast?latitude=21.1702&longitude=81.2402&current=temperature_2m",
        method="GET"
    )
)

# Provide Data Model with validation rules in descriptions for each field.
class LoanFlow(ConversationalDataModel):
    loan_type: str = Field(None, description="Type of loan: personal, home")
    annual_income: int = Field(None, description="Annual income of the applicant in INR")
    credit_score: int = Field(None, description="Credit score of the applicant. Must be between 300 and 850")
    property_value: int = Field(None, description="Value of the property for home loan in INR")
    loan_amount: int = Field(None, description="Desired loan amount in INR. MUST be greater than ₹11 lakh for approval")

loan_application = ConversationalGraph(
        name="Loan Application",
        DataModel= LoanFlow,
        off_topic_threshold=5,
        graph_type="STRICT"
    )
    
# Start Greeting
q0 = loan_application.state(
    name="Greeting",
    instruction="Welcome user and start the conversation and tell the weather",
    tool=get_weather
)

# Loan Type Selection
q1 = loan_application.state(
    name="Loan Type Selection",
    instruction="Ask user to select loan type. We only offer personal loan and home loan at the moment.",
)

# Personal Loan Details
q1a = loan_application.state(
    name="Personal Loan Details",
    instruction="Collect annual income and credit score for personal loan. Annual income must be > ₹5 lakh and credit score > 500."
)

# Home Loan Details
q1b = loan_application.state(
    name="Home Loan Details",
    instruction="Collect annual income, credit score, and property value for home loan. Annual income must be > ₹6 lakh, credit score > 650, and property value must be reasonable."
)

# Merged State → Loan Amount Collection
q2 = loan_application.state(
    name="Loan Amount Collection",
    instruction="Collect the desired loan amount. Minimum loan amount is ₹11 lakh for approval."
)


q2a = loan_application.state(
    name="Loan Amount Rejection",
    instruction="Inform user that the requested loan amount is below the minimum requirement of ₹11 lakh and politely end the application process."
)   

q2b = loan_application.state(
    name="Loan Amount Acceptance",
    instruction="Proceed to the next step of the application process as the requested loan amount meets the minimum requirement."
)

q3 = loan_application.state(
    name="Loan Application Complete",
    instruction="Inform user that the application has been submitted successfully"
)

# Master / Off-topic handler
q_master = loan_application.state(
    name="Off-topic Handler",
    instruction="Handle off-topic or inappropriate inputs respectfully and end the call politely",
    master=True
)


# Define Transitions

# Greeting → Loan Type Selection
loan_application.transition(
    from_state=q0,
    to_state=q1,
    condition="User ready to apply for loan"
)

# Branch from Loan Type Selection
loan_application.transition(
    from_state=q1,
    to_state=q1a,
    condition="User wants personal loan"
)

loan_application.transition(
    from_state=q1,
    to_state=q1b,
    condition="User wants home loan"
)

# Merge all branches → Loan Amount Collection
loan_application.transition(
    from_state=q1a,
    to_state=q2,
    condition="Personal loan details collected and verified"
)

loan_application.transition(
    from_state=q1b,
    to_state=q2,
    condition="Home loan details collected and verified"
)

# Loan Amount → Review and Confirm
loan_application.transition(
    from_state=q2,
    to_state=q2a,
    condition="Loan application rejected due to insufficient amount (< ₹11 lakh)"
)

loan_application.transition(
    from_state=q2,
    to_state=q2b,
    condition="Loan amount meets minimum requirement (≥ ₹11 lakh)"
)

# Review → Complete
loan_application.transition(
    from_state=q2b,
    to_state=q3,
    condition="User confirms submission of loan application"
)

Core Concepts

ConversationalGraph

The main entry point. It holds all your states and transitions.

• name: Name of your conversational graph.
• DataModel: A Pydantic model class that defines the schema for data you want to collect during the conversation.
• off_topic_threshold: Number of times the conversation can go off-topic before the flow jumps to the master state.
• graph_type: Currently supports "STRICT" for rigid flows.

State

A node in the conversation graph.

• name: Name of the state.
• instruction: Instructions for the LLM on what to do in this state.
• tool: A Python function to execute in this state (optional).
• master: If True, this state acts as a master state (optional).
• required_fields: List of field names (matching your DataModel) that must be collected or present before the tool can run or the state is considered 'complete' (optional).

Transition

Defines the valid paths between states.

• from_state: Source state name.
• to_state: Destination state name.
• condition: Description of when this transition should happen. The LLM uses this to decide the next step.

Edge Cases

• There can be zero or one master state.
• Tools must be accessible.
• Every transition must point to an existing state.
• There can only be one start state.