OmniRetrieval Refuses to Embed Everything
OmniRetrieval (arXiv 2605.29250, 61 upvotes on HF) from KAIST takes the opposite stance of every RAG stack in production. Instead of crunching tables, knowledge graphs, and text into one shared embedding space, it routes the natural-language query to the right source and dispatches a source-NATIVE query β SQL to the database, SPARQL or graph traversal to the KG, semantic search to text. The retriever stops being a vector store and starts being a dispatcher.
Tested across 13 datasets and 309 knowledge bases covering text, relational, and graph sources, it beats single-source baselines. The author list runs Jinheon Baek, Sung Ju Hwang, and six collaborators β the same KAIST group that has been steadily pushing on retrieval+LLM for a couple years now.
The spine: agents need to query MANY sources, and the lazy answer is to homogenize. Embed everything, hope cosine similarity does the right thing, accept that joins are now soft. The actual answer is that databases were already good at being databases. Tables know how to be queried as tables. The win is teaching the LLM which surface to talk to and how, not forcing every surface to look like a vector store.
This is what tool use should look like once you take it seriously. Right now most agents see a knowledge base as 'retrieve top-K chunks.' That is throwing away the structure the source already has. OmniRetrieval is the early move toward agents that pick the right query language for the right shape of knowledge, the way a human analyst does.
Paper: https://arxiv.org/abs/2605.29250
← Back to all articles
Tested across 13 datasets and 309 knowledge bases covering text, relational, and graph sources, it beats single-source baselines. The author list runs Jinheon Baek, Sung Ju Hwang, and six collaborators β the same KAIST group that has been steadily pushing on retrieval+LLM for a couple years now.
The spine: agents need to query MANY sources, and the lazy answer is to homogenize. Embed everything, hope cosine similarity does the right thing, accept that joins are now soft. The actual answer is that databases were already good at being databases. Tables know how to be queried as tables. The win is teaching the LLM which surface to talk to and how, not forcing every surface to look like a vector store.
This is what tool use should look like once you take it seriously. Right now most agents see a knowledge base as 'retrieve top-K chunks.' That is throwing away the structure the source already has. OmniRetrieval is the early move toward agents that pick the right query language for the right shape of knowledge, the way a human analyst does.
Paper: https://arxiv.org/abs/2605.29250
Comments