Perplexity Answer Engine Architecture: Retrieval, Grounded Generation and Citations

As a product manager at Perplexity, how would you design the system behind its answer engine? ```hint Follow one query end to end Trace a single question from the moment it is typed until a cited answer is on screen. At every step, ask what it adds in time and cost, and how it could make the answer wrong. ``` ### Clarifying Questions - Should I go deep on the architecture (retrieval, ranking, generation, serving), or stay at the level of components and the product trade-offs between them? - Is the scope quick answers to everyday questions, slower multi-step research, or both? - Should the system run its own crawler and web index, call third-party search APIs, or combine the two? - Which constraint should drive the design: time until the answer starts appearing, cost per query, or accuracy of the citations? ### What a Strong Answer Covers - The end-to-end query path: understanding the question, retrieving and ranking sources, generating an answer from them, and attaching citations. - How the system keeps answers grounded in what the sources say, and how citation accuracy is checked. - Latency and cost trade-offs across retrieval depth, model choice, caching and streaming, including how a quick mode and a deep research mode differ. - Index freshness for time-sensitive questions, and the build-versus-buy choice for search. - Quality metrics and failure handling, including unsupported citations, low-quality sources and instructions planted in web pages. ### Follow-up Questions - Users report that some citations do not support the sentence they are attached to. How would you detect this at scale and reduce it? - Moving to a stronger model doubles the cost per query. Which levers would you try before rolling the change back? - How would the design change to let a research task run for several minutes and read dozens of sources?