LLM pipelines and serverless AI workflows

// Stage 1 — extract-text.mjs
// Input: { documentUrl }
// Output: { text, charCount }
export async function handler(event) {
  const { documentUrl } = event.payload ?? JSON.parse(event.body || '{}');
  const text = await extractText(documentUrl); // PDF/HTML/DOCX → plain text
  return { text, charCount: text.length };
}

// Stage 2 — classify.mjs
// Input: previousOutput.text
// Output: { text, category } — 'invoice' | 'contract' | 'report'
export async function handler(event) {
  const { text } = event.previousOutput;
  const category = await classifyWithLLM(text);
  return { ...event.previousOutput, category };
}

// Stage 3 — retrieve.mjs
// Retrying this stage does NOT re-run moderation or classification
export async function handler(event) {
  const { text, category } = event.previousOutput;
  const related = await vectorSearch(text, { filter: { category } });
  return { ...event.previousOutput, related };
}

// Stage 4 — summarize.mjs
export async function handler(event) {
  const { text, related } = event.previousOutput;
  const summary = await summarizeWithLLM(text, related);
  const tokens = summary.usage.total_tokens;
  return { summary: summary.text, tokens }; // tokens tracked per stage
}

// Stage 5 — store-and-notify.mjs
export async function handler(event) {
  const { summary, tokens } = event.previousOutput;
  const doc = await db.documents.create({ summary, tokens });
  await notify(doc.id);
  return { docId: doc.id };
}