A maintenance assistant for technicians must answer from the latest service bulletins and also explain a photo of a worn gearbox the technician uploads. Which architecture is the best fit?
Show answer and explanation
Answer: A
The photo needs a model that can interpret images, and the bulletins need answers grounded in current indexed content. A multimodal model combined with retrieval covers both needs in one response. A text-only model ignores the photo, search alone cannot explain the image or compose an answer, and OCR reads text but not the physical condition of the gearbox.
Why the other options are wrong
- B. Fine-tuning fixes the model's knowledge at training time, so bulletins published later would be missing. Discarding the photo also loses the evidence the technician needs.
- C. Search returns matching documents but cannot interpret a gearbox photo or compose a grounded answer. Some generative component is needed for the explanation.
- D. OCR only extracts visible text from the photo. It cannot assess wear and does not retrieve current bulletins from an index.