Perplexity Deep Research: Search as Code and the WANDR Benchmark
Perplexity has fundamentally shifted the technical architecture of its autonomous research capabilities from an iterative query-results loop to a paradigm known as Search as Code (SaC). This shift drastically improves efficiency on complex, multi-step "wide research" tasks, as measured by Perplexity's own WANDR benchmark and the industry-standard open-source DRACO benchmark.
The "Search as Code" (SaC) Paradigm
Traditional agentic search relies on an iterative loop: the agent generates a query, calls a search tool, receives a snippet list, reads the context, generates a new query, and repeats. This approach is highly inefficient, consuming massive token context and introducing significant latency due to repeated LLM inference steps.
With Search as Code (SaC), Perplexity replaces this conversational loop with programmatic execution:
- Code Generation: The model writes a Python script designed to execute a complex, multi-stage search pipeline.
- Sandboxed Execution: The script runs in an isolated sandbox environment. It can fan out multiple search queries in parallel, filter and deduplicate results programmatically, and perform calculations.
- Context Optimization: Instead of dumping thousands of raw search result tokens into the LLM context, the sandboxed script processes the data locally and hands back only the highly relevant, distilled synthesis.
On Hacker News, practitioners have noted the advantages of this approach:
"This architecture turns that into one program: fan out the greps, filter and dedupe in the sandbox, hand back only what matters... So it's less about replacing sequential exploration and more about making each grep step way wider without flooding context" — 1zael on Hacker News
However, some point out that this introduces new customer support and optimization challenges:
"The generated Python code is unlikely to always run queries in an optimal way. Normally you'd tell your customer they need to fix their code if they're doing something inefficient, but here, they didn't write the code and are presumably not expected to understand it." — nitwit005 on Hacker News
Evaluating Deep Research: WANDR and DRACO Benchmarks
To justify and measure this architectural shift, Perplexity introduced two major benchmarks:
1. The WANDR (Wide Research) Benchmark
WANDR is an in-house benchmark designed specifically to mirror real-world, professional "wide research" tasks that require broad information-seeking across multiple domains. On WANDR, Search as Code achieves a score of 0.386, outperforming traditional iterative search loops by up to 2.5x while cutting token consumption by up to 85% on long-horizon tasks.
2. The DRACO Benchmark
The Deep Research Accuracy, Completeness, and Objectivity (DRACO) benchmark is an open-source cross-domain benchmark published by Perplexity on Hugging Face. It consists of complex, open-ended research tasks evaluated against expert-curated rubrics.
- Unlike typical academic benchmarks, DRACO evaluates agents equipped with a full research toolset, including code execution and browser capabilities.
- On DRACO, Perplexity Deep Research (utilizing Claude Opus 4.6) achieves 70.5% accuracy, highlighting the strength of combining frontier reasoning models with programmatic search pipelines.