Public: Perplexity answer engine and deep research stack

Architectural notes

Per the Vespa.ai engineering blog and an April 2025 partnership announcement,
Perplexity brought web search fully in-house on Vespa: one engine combining
vector search, lexical search, structured filtering and ML ranking over an
index of billions of web pages and attached user files, answering in roughly
a hundred milliseconds at thousands of queries per second. The same post
describes the RAG loop (query analysis, retrieval, LLM generation) iterating
multiple times for deep research.

In-house models do not map to the catalog: Sonar is a Llama 3.3 70B
fine-tune served at ~1,200 tokens/sec on Cerebras inference hardware (per
the Meet New Sonar post), and R1-1776 is a post-trained DeepSeek-R1
reasoning model with weights on Hugging Face, served via the Sonar API.
Pro users select among Sonar and frontier models, hence model:
user-configurable. NOT publicly documented: embedding models, crawler
design, eval/observability stack, and orchestration internals.