Enterprise Knowledge Search
Perplexity needed to customize frontier models for a search experience that returns direct answers instead of links, but rapid model releases, high query volume, and the need for low latency and high throughput made continuous post-training and retrieval optimization a core challenge.
Reported impact
Perplexity processed 340 million queries a month, integrated into more than 1,500 organizations, and used NeMo-based tuning to produce Sonar models that improved search quality by 20% over newly released base models within days while preserving the low latency users expect.
Perplexity used NVIDIA NeMo to create and refresh its Sonar family of online language models for the company's answer engine. The team relies on NeMo to scale supervised fine-tuning, direct preference optimization, and proximal policy optimization across a range of model sizes and architectures, including Llama and Mistral families. Retrieval-augmented generation connects those tuned models to live source material so answers stay current and grounded, while NeMo's distributed training support lets Perplexity experiment quickly after new open-source models appear. The same stack is also used to fine-tune open-source embedding models that improve the retrieval layer itself. This architecture lets Perplexity update both generation and retrieval in a matter of days, rather than leaving a web-scale search product tied to static generic models that are slower and less factual.
Search intent and retrieved web context flow into Sonar models that were post-trained in NeMo with SFT, DPO, and PPO; tuned embedding models improve retrieval quality; and the resulting RAG pipeline returns concise grounded answers to users and enterprise integrations.
As of October 2024, the platform processes an impressive 340 million queries per month.
NeMo allows Perplexity to quickly finetune a variety of open-source embedding models.
Perplexity has applied fine-tuning to frontier models, including Llama and Mistral model families, and is leveraging retrieval-augmented generation.
Within a few days of a new open-source release, the team had a new Sonar model that was 20% improved over the base model on search.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
The public company page includes numberOfEmployees value 1408 and a PostalAddress with streetAddress 115 Sansome St, addressLocality San Francisco, and addressCountry US.
Related
NVIDIA Sample Jul 7, 2026 Siam AI Corporation Co., Ltd. (Siam.AI Cloud) Thailand tourism information was fragmented across multiple websites, difficult to verify, often outdated, and hard for visitors to access across languages, creating poor experiences and inefficient support workflows. NVIDIA AI Enterprise · NVIDIA NIM · NVIDIA NeMo NVIDIA Sample Feb 2, 2026 NVIDIA NVIDIA needed to move employee-built chatbots and agents out of shadow-AI pilots and onto a unified, secure, high-performance platform that could support hundreds of governed AI workflows across engineering, software… from 15 hours to one hour NVIDIA AI Blueprint for Retrieval-Augmented Generation (RAG) · NVIDIA AI Enterprise · NVIDIA GPU Operator NVIDIA Sample Est. · Dec 8, 2024 Cohesity Cohesity needed to improve retrieval accuracy in its generative AI assistant so relevant enterprise documents were not missed before answer generation, especially in high-stakes use cases such as legal eDiscovery. Gaia’s errors were reduced by 13%. CUDA Python · NVIDIA A100 Tensor Core GPUs · NVIDIA NIM microservices 