Research & insights
Learned Contextual Chunking: From LLM Supervision to Fast CPU Inference
How Valyu uses LLM supervision to train XGBoost document chunkers on CPU, comparing evidence retrieval and chunking latency against Gemma 12B.
The Valyu blog
Practical guides, research, and ideas from the team building search for AI.
DeepResearch is an AI-native research API. SoTA performance, at less than half the cost of the next-best system. Built for enterprise knowledge work and frontier AI applications.
Read the storyHow Valyu uses LLM supervision to train XGBoost document chunkers on CPU, comparing evidence retrieval and chunking latency against Gemma 12B.
How a real-time horizon-scanning capability powered by Valyu is helping the Nuclear Decommissioning Authority bring relevant developments into view sooner.
How Valyu serves a five-model routing stack from one Blackwell GPU with CUDA MPS, affinity routing, task-specific batching, and CRIU snapshot restore.
How o11 (YC W26) uses Valyu to ground its financial AI agents in SEC filings, company data, news, and more - skipping months of indexing and partnerships.
ChemRxiv chemistry preprints are now in Valyu - full-text search for AI agents, across the Search and DeepResearch APIs.
How Storke uses Valyu APIs to cut clinical evaluation timelines by half, grounding clinical evaluation reports in academic and regulatory literature.
How Valyu rebuilt vector index initialisation from graph walks to CPU SIMD sweeps to GPU brute force, cutting 100M-vector builds to about 25 minutes on a single GPU.
Look inside the buy-side workflow we saw in 4.2 million queries: 300 tickers a week, run by two people.