Class MultimodalAggregator

Adds auto-generated captions to ExtractedImage objects that lack one, using a caller-supplied vision LLM function.

Images are processed in parallel via Promise.allSettled() so a single failed captioning attempt does not block the rest. Images whose captioning fails retain their original (un-captioned) state rather than propagating the error.

Example — with a vision LLM

const aggregator = new MultimodalAggregator({
  describeImage: async (buf, mime) => myVisionLLM.describe(buf, mime),
});

const captioned = await aggregator.processImages(doc.images ?? []);
Copy

Example — passthrough (no LLM configured)

const aggregator = new MultimodalAggregator();
const unchanged  = await aggregator.processImages(doc.images ?? []);
Copy