Generate more successful research trajectories without wasting context, compute, or engineering time. Tavily gives post-training teams high-quality web evidence, configurable retrieval, and infrastructure that scales to thousands of concurrent rollouts.
NVIDIA generated ~80,000 research trajectories using Tavily to fine-tune Nemotron 3 Super. The model helps power NVIDIA's AI-Q deep researcher, which achieved #1 on DeepResearch Bench I and II.
Read the story
Retrieval quality that improves task success
Context efficiency that reduces inference waste
Reliability and throughput built for large rollout volumes
Hands-on optimization for long-horizon agent harnesses
#1
on DeepResearch Bench I and II
NVIDIA
NVIDIA used Tavily to generate ~80,000 research trajectories, with ~67,000 retained after filtering to fine-tune Nemotron 3 Super.
100 million
credits consumed in 6 weeks
Top AI research lab
A leading AI research lab uses Tavily to power live web search during reinforcement learning and generate trajectories for supervised fine-tuning.
30-40
step research trajectories
AI R&D Company
Across 30–40 step research trajectories, Tavily worked with the lab to optimize retrieval configuration, memory, and context management to improve long-horizon performance.
Give agents structured, source-backed, information-dense results that help them find usable evidence across complex, multi-step research trajectories.
Control result count, content length, search depth, and domains to reduce unnecessary tokens while preserving the evidence agents need.
Run repeated, concurrent agent calls with up to 1,000 searches per second, backed by enterprise SLAs, SOC 2 Type II, and ISO 27001 certification.
Work directly with Tavily on retrieval configuration, fetching, memory, context management, and long-horizon agent behavior.
/partners
Benefit from our global partner network
Post-training improves an AI model's capabilities and behavior after pretraining. Common methods include supervised fine-tuning (SFT), which trains models on examples of desired behavior, and reinforcement learning (RL), which uses rewards to guide model weight updates.
For large language models (LLMs) used in research agents, post-training can improve multi-step reasoning, web search, tool use, and evidence synthesis.
Web search gives agents access to external evidence while they complete tasks during post-training. Their training trajectories record the sequence of model actions, search queries, tool responses, and answers.
For supervised fine-tuning, teams select high-quality trajectories as training examples. For reinforcement learning, teams score task outcomes and use rewards to update model weights. Tavily supplies live web results within these workflows; the training pipeline handles trajectory collection, evaluation, and learning.
Tavily provides the web retrieval layer inside these training workflows.
Evaluate a web search API inside your agent harness using representative training tasks. Measure task success, usable trajectories, token consumption, retrieval latency, and reliability at your expected request volume. Compare total cost per successful rollout, including search and model inference.
Tavily lets teams configure search depth, result count, content length, and domain filters to evaluate how retrieval settings affect training outcomes.
Tavily provides information-dense, structured web results that agents can use as evidence during training. Relevant content snippets and source URLs help models gather information while giving research teams visibility into the evidence behind each trajectory.
Teams can adjust search depth to balance relevance and latency, limit returned content to manage context usage, and filter domains to control sources. Tavily also works with research teams to optimize retrieval within their training and evaluation harnesses.
