Zacharias-Brohn
135 lines
3.4 KiB
TypeScript
135 lines
3.4 KiB
TypeScript
'use server';
|
|
|
|
import ollama, { Tool } from 'ollama';
|
|
|
|
export interface ChatMessage {
|
|
role: 'user' | 'assistant' | 'system' | 'tool';
|
|
content: string;
|
|
tool_calls?: any[]; // Ollama tool calls
|
|
images?: string[];
|
|
}
|
|
|
|
// --- Tool Definitions ---
|
|
|
|
const tools: Tool[] = [
|
|
{
|
|
type: 'function',
|
|
function: {
|
|
name: 'get_current_time',
|
|
description: 'Get the current time',
|
|
parameters: {
|
|
type: 'object',
|
|
properties: {},
|
|
required: [],
|
|
},
|
|
},
|
|
},
|
|
{
|
|
type: 'function',
|
|
function: {
|
|
name: 'calculate',
|
|
description: 'Perform a mathematical calculation',
|
|
parameters: {
|
|
type: 'object',
|
|
properties: {
|
|
expression: {
|
|
type: 'string',
|
|
description: 'The mathematical expression to evaluate (e.g., "2 + 2" or "15 * 7")',
|
|
},
|
|
},
|
|
required: ['expression'],
|
|
},
|
|
},
|
|
},
|
|
];
|
|
|
|
// --- Tool Implementations ---
|
|
|
|
const availableTools: Record<string, Function> = {
|
|
get_current_time: () => {
|
|
return new Date().toLocaleTimeString();
|
|
},
|
|
calculate: ({ expression }: { expression: string }) => {
|
|
try {
|
|
// Safety: simplistic eval for demo purposes.
|
|
// In production, use a math parser library like mathjs.
|
|
// eslint-disable-next-line no-eval
|
|
return eval(expression).toString();
|
|
} catch {
|
|
return 'Error evaluating expression';
|
|
}
|
|
},
|
|
};
|
|
|
|
export async function chat(model: string, messages: ChatMessage[]) {
|
|
try {
|
|
// 1. Initial Call
|
|
let response;
|
|
try {
|
|
response = await ollama.chat({
|
|
model: model,
|
|
messages: messages,
|
|
tools: tools,
|
|
});
|
|
} catch (e: any) {
|
|
// Fallback: If model doesn't support tools, retry without them
|
|
if (e.message?.includes('does not support tools')) {
|
|
console.warn(`Model ${model} does not support tools. Falling back to standard chat.`);
|
|
response = await ollama.chat({
|
|
model: model,
|
|
messages: messages,
|
|
});
|
|
} else {
|
|
throw e;
|
|
}
|
|
}
|
|
|
|
// 2. Loop to handle tool calls (Ollama might chain multiple calls)
|
|
// We limit recursion to avoid infinite loops
|
|
let maxTurns = 5;
|
|
|
|
while (response.message.tool_calls && response.message.tool_calls.length > 0 && maxTurns > 0) {
|
|
maxTurns--;
|
|
|
|
// Append the assistant's message (which contains the tool calls) to history
|
|
messages.push(response.message as ChatMessage);
|
|
|
|
// Execute each tool call
|
|
for (const tool of response.message.tool_calls) {
|
|
const functionName = tool.function.name;
|
|
const functionToCall = availableTools[functionName];
|
|
|
|
if (functionToCall) {
|
|
console.log(`🤖 Tool Call: ${functionName}`, tool.function.arguments);
|
|
const functionArgs = tool.function.arguments;
|
|
const functionResponse = functionToCall(functionArgs);
|
|
|
|
// Append the tool result to history
|
|
messages.push({
|
|
role: 'tool',
|
|
content: functionResponse,
|
|
});
|
|
}
|
|
}
|
|
|
|
// 3. Send the tool results back to the model to get the final answer
|
|
response = await ollama.chat({
|
|
model: model,
|
|
messages: messages,
|
|
tools: tools,
|
|
});
|
|
}
|
|
|
|
return {
|
|
success: true,
|
|
message: response.message,
|
|
};
|
|
} catch (error: any) {
|
|
console.error('Chat error:', error);
|
|
return {
|
|
success: false,
|
|
error: error.message || 'Failed to generate response',
|
|
};
|
|
}
|
|
}
|