Valyu AI
Valyu sells a retrieval API positioned as the data layer for research agents: one interface to search specialist, proprietary, and web sources and run multi-step research that returns cited, structured output. Its stated focus sectors are finance and life sciences / R&D, covering specialist scientific sources such as preprint archives alongside general web results.
The pitch is evidence control — retrieval over curated primary sources with citations attached, rather than whatever a general crawl surfaces.
Website: https://www.valyu.ai
The API, as Used in Practice
From the worked example in How to Use Jev - A Practical Guide, the search call takes an
included_sources allowlist (e.g. PubMed and arXiv collections), a start_date, a
max_num_results, and a relevance_threshold — so filtering happens at retrieval time rather
than downstream. Results carry title, URL, and content.
That threshold parameter is the interesting part for this vault: it presumes a comparable relevance score at the retrieval stage, which is the same property calibrated scoring provides at the rerank stage.
Relevance to Search
Valyu is a federated retrieval product over curated source collections — the “fetch precisely” half of the retrieve-then-judge pattern its own guide describes, where a cheap per-passage judgment (Reranking) filters the retrieved set before anything expensive consumes it. The argument it makes for itself is a grounding argument rather than a ranking one: a downstream model’s judgment can only be as good as the corpus it was handed.
In This Vault
- How to Use Jev - A Practical Guide — published on Valyu’s dev.to account by Prosper Otemuyiwa; a guide to Jev whose Pattern 5 pairs Valyu retrieval with per-passage Noul filtering. Vendor-adjacent, and the product placement should be read as such, though the underlying observation about grounding stands on its own.
Related Concepts
- Reranking — the judging half of the pattern
- Federated Search · Search Scopes — searching across curated source collections
- Calibrated Relevance Probability — comparable scores as a precondition for thresholding
- RAG · Clean Context · Context Engineering — what precise retrieval is in service of
- Semantic Relevance — relevance thresholding at retrieval time
Related Notes
- TypeSafe · Jev — the judging model in the pattern
- How to Use Jev - A Practical Guide — the source