kuluru vineeth
← builds

Self-hosted search

Search that decides relevance by holding an election.

OpenBeam is open-source search across everything your team already uses. It reads your tools, and when you ask it something, it does not trust any single way of finding the answer.

your toolsslackgmaildrivenotiongithubindexhybridevery lanereads it"q3 pricing deck"one query, two formsbm25exact wordsdensemeaningsparserare tokensone ranked answer
87 connectorshybrid retrievalMCP serverself-hostedAGPL

The idea it taught me

A search for "q3 pricing deck" is three questions at once. There may be a file named exactly that, which rewards matching the words. The right document might be called "autumn revenue review" and never say deck, which rewards matching the meaning. And "q3" carries more weight than its three letters suggest, which rewards a model that has learned which rare words matter.

OpenBeam answers all three separately, then holds an election. Each retriever ranks the documents its own way, and instead of trying to add up scores that live on different scales, it keeps only the ranks and lets them vote. A document no single retriever loved but all three liked can win, which is usually the right call.

Read the full mechanism →

Go deeper

EssayThree retrievers, one answerThe full pipeline: two representations of a query, three lanes, rank fusion, and the funnel that spends compute where it is cheap to be wrong.BuildThe OpenBeam seriesThe whole thing built out, on the channel.Codegithub.com/kuluruvineeth/openbeamSelf-hostable, 87 connectors, an MCP server. Read the ranking in packages/services.