Synthesis generates a novel answer by combining evidence scattered across multiple retrieved documents. It addresses cases where no single passage contains the complete answer and extraction from the strongest passage misses relevant context. section 5.0 section 5.7
A practical synthesis architecture separates evidence isolation from evidence combination:
- Encode each passage independently and in parallel, preserving separate representations and allowing processing across many documents. section 5.0 section 5.7
- Fuse the passage representations in the decoder, which attends to all retrieved evidence while generating each answer step. section 5.7
- Prefer retrieval recall over perfect ranking: relevant information only needs to appear somewhere in the top-k set, although larger sets increase decoding latency. section 5.7
Apple revenue growth
A financial analyst asks about Apple's revenue growth using passages on emerging-market iPhone sales, Services growth, and resilient logistics. A synthesizing reader should combine all three drivers rather than return only the first strong match. In practice, smaller general-purpose readers often remain extractive and let the strongest passage dominate, revealing that multi-document access alone does not guarantee synthesis. section 5.5
