ctx.llm

const response = await ctx.llm.complete({
  prompt: 'Summarize the following document in 3 bullet points.',
  temperature: 0.3,
  maxTokens: 500,
});

console.log(response.content);        // Generated text
console.log(response.cost.usd);       // e.g. 0.0023
console.log(response.usage.totalTokens); // e.g. 847
console.log(response.provenanceId);   // Audit trail reference

// Use message arrays for multi-turn conversations
const response = await ctx.llm.complete({
  system: 'You are a financial analyst. Be precise and cite sources.',
  prompt: [
    { role: 'user', content: 'What are the key risks in this contract?' },
    { role: 'assistant', content: 'I see three main areas of concern...' },
    { role: 'user', content: 'Focus on the indemnification clause.' },
  ],
  temperature: 0.2,
  maxTokens: 1000,
});

// Stream tokens in real-time
for await (const chunk of ctx.llm.stream({ prompt: userQuery })) {
  process.stdout.write(chunk.delta);  // Print new tokens as they arrive

  if (chunk.finishReason) {
    console.log('\nDone. Tokens used:', chunk.usage?.totalTokens);
  }
}

// Generate embeddings for semantic search
const embedding = await ctx.llm.embed('How do I process an invoice?');

// embedding.vector is a float array ready for vector search
const results = await ctx.memory.persistent.search(embedding.vector, {
  limit: 5,
  minScore: 0.7,
});

console.log(embedding.dimensions);  // e.g. 1536
console.log(embedding.cost.usd);    // Embedding costs are tracked too

// Track costs across multiple LLM calls
const step1 = await ctx.llm.complete({ prompt: 'Extract entities from: ' + text });
const step2 = await ctx.llm.complete({ prompt: 'Classify entities: ' + step1.content });

// Cost of the last call
const lastCost = ctx.llm.getLastCost();

// Total cost of ALL llm calls in this execution
const totalCost = ctx.llm.getTotalCost();
console.log(`Total: $${totalCost.usd.toFixed(4)}`);  // e.g. Total: $0.0047

// Real-world: invoice extraction (from packages/agents-reference)
const extractionResult = await ctx.llm.complete({
  prompt: [
    {
      role: 'user',
      content: `Extract invoice data as JSON with fields:
        vendor, date, total, line_items (array with description, qty, unit_price, total).
        Invoice content:
        ${invoiceContent}`,
    },
  ],
  temperature: 0.2,   // Low temp for structured extraction
  maxTokens: 2000,
});

const invoiceData = JSON.parse(extractionResult.content);