LastSearch
Research infrastructure for AI agents that performs real-time web search with claim-by-claim verification against source pages, returning cited answers with an evidence-based confidence score instead of a model's own self-assessment.
LastSearch gives an AI agent grounded web search: it searches the web, fetches pages, reranks results, extracts individual claims from the answer, and verifies each claim against the actual text of the source page before returning a cited answer, so every claim is backed by a real, checked source rather than an unverified citation. Compound claims are automatically split into individual atomic facts that are verified separately, and a purpose-built verification model, trained on over a million claim-evidence pairs, reranks candidate evidence and applies contradiction penalties.
Confidence scoring combines hybrid keyword and semantic matching, cross-referencing multiple independent sources, and a domain-authority ranking across more than ten thousand sites, from institutional government and education domains down through news, tech journalism, and community sources. The project is available as an MCP server, a REST API, and Python and JavaScript SDKs, including direct integrations for LangChain, CrewAI, and LlamaIndex.