sarvam-conv-ai-sdk 1.0.1

A Python SDK for building multilingual, context-aware conversational agents on Sarvam's Samvaad platform.

Sarvam Conv AI SDK

The Sarvam Conversational AI SDK is a Python package that helps developers build and extend conversational agents. It provides core components to manage conversation flow, language preferences, and messaging, making it easier to develop interactive and context-aware AI experiences.

---

Overview

The Sarvam Conv AI SDK enables developers to create tools that can:

• Facilitate agentic capabilities like API calling in the middle of a conversation.
• Manage agent-specific variables and state across interactions
• Control and modify the language used during conversations
• Send dynamic messages to both the user and the underlying language model (LLM)

---

Installation

Install the SDK via pip:

pip install sarvam-conv-ai-sdk

---

Example Usage

import httpx
from pydantic import Field
from sarvam_conv_ai_sdk import (
    SarvamToolLanguageName,
    SarvamOnEndTool,
    SarvamOnStartTool,
    SarvamTool,
    SarvamOnStartToolContext,
    SarvamToolContext,
    SarvamOnEndToolContext,
    SarvamToolOutput
)

class OnStart(SarvamOnStartTool):
    async def run(self, context: SarvamOnStartToolContext):
        user_id = context.get_user_identifier()
        async with httpx.AsyncClient() as client:
            response = await client.get(f"https://sarvam-flights.com/users/{user_id}")
            response.raise_for_status()
            user_data = response.json()

        source_destination = user_data.get("home_city")
        context.set_agent_variable("source_destination", source_destination)
        context.set_agent_variable("passenger_name", user_data.get("name"))
        context.set_initial_language_name(SarvamToolLanguageName.ENGLISH)
        context.set_initial_bot_message(
            f"Hello! Would you like to book a flight from {source_destination}? Where would you like to go?",
        )
        return context


class BookFlight(SarvamTool):
    """Book a flight based on the user's travel preferences."""

    destination: str = Field(description="City of destination")
    travel_date: str = Field(description="Date of travel (YYYY-MM-DD)")

    async def run(self, context: SarvamToolContext) -> SarvamToolOutput:
        source_destination = context.get_agent_variable("source_destination")
        booking_data = {
            "source": source_destination,
            "destination": self.destination,
            "travel_date": self.travel_date,
            "passenger_name": context.get_agent_variable("passenger_name"),
        }

        async with httpx.AsyncClient() as client:
            response = await client.post(
                "https://sarvam-flights.com/book", json=booking_data
            )
            response.raise_for_status()
            booking_result = response.json()

        if booking_result.get("status") == "confirmed":
            context.set_agent_variable("booking_id", booking_result.get("booking_id"))
            context.set_end_conversation()
            return SarvamToolOutput(
                message_to_user=f"Flight booked successfully to {self.destination}!",
                context=context
            )
        else:
            return SarvamToolOutput(
                message_to_llm="Booking failed. Please suggest similar destinations.",
                context=context
            )


class OnEnd(SarvamOnEndTool):
    async def run(self, context: SarvamOnEndToolContext):
        feedback = context.get_agent_variable("feedback")
        negative_words = ["bad", "poor", "disappointed", "unhappy", "problem"]
        is_negative = any(word in feedback.lower() for word in negative_words)
        context.set_agent_variable("feedback_sentiment", is_negative)
        return context

---

Base Classes

The SDK exposes three base classes for tool development:

1. SarvamTool

Primary base class for all operational tools invoked during conversation flow.

Example:

class MyCustomTool(SarvamTool):
    """Brief description of the tool's purpose."""

    tool_variable: type = Field(description="Description of this input parameter")

    async def run(self, context: SarvamToolContext) -> SarvamToolOutput:
        # Custom tool logic
        return SarvamToolOutput(
            message_to_user="Response to user",
            message_to_llm="Context for LLM",
            context=context
        )

2. SarvamOnStartTool

Executed at the beginning of a conversation, typically for initialization. The class must be named OnStart.

3. SarvamOnEndTool

Executed at the end of a conversation, typically for cleanup or post-processing. The class must be named OnEnd.

---

Context Classes and Methods

SarvamToolContext

The context object passed to SarvamTool.run() methods.

Variable Management

• get_agent_variable(variable_name: str) -> Any Retrieve the value of a variable.

• set_agent_variable(variable_name: str, value: Any) -> None Update a variable's value.

Language Control

• get_current_language() -> SarvamToolLanguageName Returns the current language preference.

• change_language(language: SarvamToolLanguageName) -> None Update the language preference.

Conversation Flow

• set_end_conversation() -> None Explicitly end the conversation.

---

SarvamOnStartToolContext

The context object passed to SarvamOnStartTool.run() methods.

Variable Management

• get_agent_variable(variable_name: str) -> Any Retrieve the value of a variable.

• set_agent_variable(variable_name: str, value: Any) -> None Update a variable's value.

User Information

• get_user_identifier() -> str Get the user identifier.

Initialization Methods

• set_initial_bot_message(message: str) -> None Set the first message sent by the agent when the conversation starts.

• set_initial_state_name(state_name: str) -> None Set the initial state from which the agent should start.

• set_initial_language_name(language: SarvamToolLanguageName) -> None Define the initial language preference for the user.

---

SarvamOnEndToolContext

The context object passed to SarvamOnEndTool.run() methods.

Variable Management

• get_agent_variable(variable_name: str) -> Any Retrieve the value of a variable.

• set_agent_variable(variable_name: str, value: Any) -> None Update a variable's value.

User Information

• get_user_identifier() -> str Get the user identifier.

Interaction Transcript

• get_interaction_transcript() -> Optional[InteractionTranscript] Retrieve the conversation history containing user and agent messages in English.

Example transcript:

[
    InteractionTurn(role=<InteractionTurnRole.AGENT: 'agent'>, en_text='Hello! How can I help you today?'),
    InteractionTurn(role=<InteractionTurnRole.USER: 'user'>, en_text='I need to book a flight'),
    InteractionTurn(role=<InteractionTurnRole.AGENT: 'agent'>, en_text='I can help you with that. Where would you like to go?'),
    InteractionTurn(role=<InteractionTurnRole.USER: 'user'>, en_text='I want to go to Mumbai'),
    InteractionTurn(role=<InteractionTurnRole.AGENT: 'agent'>, en_text='Great! When would you like to travel?')
]

---

Return Types

SarvamToolOutput

The return type for SarvamTool.run() methods. Contains:

• message_to_user: Optional[str] - Message that is sent directly to the user
• message_to_llm: Optional[str] - Message that is sent to the LLM, which then responds
• context: SarvamToolContext - The updated context object

Note: At least one of message_to_llm or message_to_user must be set.

Important: When both message_to_user and message_to_llm are set, only the message_to_user is actually sent to the user, but the message_to_llm overrides the message_to_user when adding to the chat thread for the LLM's context.

---

Supported Languages

The SDK supports multilingual conversations using the SarvamToolLanguageName enum. Available languages include:

• Bengali
• Gujarati
• Kannada
• Malayalam
• Tamil
• Telugu
• Punjabi
• Odia
• Marathi
• Hindi
• English

---

Best Practices

1. Always implement run(): The run() method is the entry point for tool execution logic.
2. Use Field() for parameters: Ensures type safety and adds descriptive metadata necessary for LLM to use in the prompt.
3. Gracefully handle errors: Avoid accessing unset variables or using invalid types.
4. Return the appropriate type: SarvamTool.run() must return SarvamToolOutput, while SarvamOnStartTool.run() and SarvamOnEndTool.run() return their respective context objects.
5. Write meaningful docstrings: Clearly describe what each tool is intended to do as this directly impacts the performance of tool calling capabilities of the agent.
6. Use async operations for I/O: For the best performance, use async/await for external API calls to avoid blocking.
7. Use context methods: Use the provided context methods for variable management, language control, and messaging instead of directly accessing context attributes.

---

Error Handling

The SDK includes built-in error handling for common scenarios:

• Variable not found: Raises ValueError when accessing undefined variables
• Variable not defined: Raises ValueError when setting variables that haven't been initialized
• Non-serializable values: Raises ValueError when variable values cannot be JSON serialized
• Invalid output: Raises ValueError when SarvamToolOutput is created without at least one message