Sarvam AI: add tool calling support

[alfredjimmy-redbus]

Tool Calling Implementation – Sarvam AI Contrib

Overview

Tool calling allows Sarvam AI to work with the ADK’s function tools. The model receives tool definitions and can return tool calls; the runner executes tools and sends results back.

---

Files Edited

1. ChatMessage.java

Role: Support tool calls and tool responses in messages.

Field	Type	Use case
tool_calls	List<ChatToolCall>	Assistant messages with function calls
tool_call_id	String	Tool messages linked to a specific call

Add getters/setters for both fields.

---

2. ChatToolCall.java

Role: Map Sarvam’s tool call format (OpenAI-style).

Structure: id, type="function", function: { name, arguments } (JSON string).

---

3. ChatRequest.java

Role: Convert ADK LlmRequest → Sarvam chat request, including tools and prior tool use.

Edit	Description
tools field	List of tools in Sarvam format
tool_choice field	Set to "auto" when tools are present
fromLlmRequest	Map Part.functionCall → assistant message with tool_calls; Part.functionResponse → tool message with tool_call_id
buildTools()	Convert FunctionDeclaration → { type, function: { name, description, parameters } }
schemaToMap()	Convert Schema to Map for JSON
schemaTypeToString()	Map Type enum → string ("string", "number", etc.)
getTools(), getToolChoice()	Getters for JSON serialization
Imports	Add Schema, Type from com.google.genai.types

---

4. SarvamAi.java

Role: Convert Sarvam responses → ADK LlmResponse.

Edit	Description
toLlmResponse()	Support both text and tool-call responses
Tool-call branch	If effectiveMsg.getToolCalls() is non-empty → parse each ChatToolCall → create FunctionCall.builder().id().name().args() → build Part.functionCall → set Content parts → return LlmResponse
Text branch	If no tool calls → use Part.fromText(content) as before
Imports	Add ChatToolCall, FunctionCall, ImmutableList, TypeReference, ArrayList, List, Map

---

Data Flow

1. Request: LlmRequest with tools and history → ChatRequest.fromLlmRequest() → request JSON.
2. Response (text): Sarvam message.content → toLlmResponse() → Part.fromText() → LlmResponse.
3. Response (tool call): Sarvam message.tool_calls → toLlmResponse() → Part.functionCall() → LlmResponse for the ADK runner.