A new research framework called LENS answers questions over documents without ever building a search index first.
Described in a paper posted August 18, LENS - short for Latent Evidence Exploration and Search - is built for cases where files change before anyone gets around to preprocessing them, and the relevant evidence depends on the specific question being asked. Instead of pre-chunking documents into embeddings or a fixed index, it keeps a running, query-specific belief about which parts of a document collection are worth checking, using a mix of lexical, local, and exploratory search moves, then asks an LLM to score relevance and narrow the search under a set budget. Evidence gets consolidated into compact, source-grounded regions and reused across related follow-up questions.
The tradeoffs show up in the numbers. On a 500-question controlled test, LENS hit 62.4% exact match and 84.8% evidence recall, against 65.2% exact match but only 50.4% evidence recall for a ReAct-style baseline that also searches without an index. On a separate 150-question test run directly over a raw Wikipedia dump with zero indexing, LENS and ReAct came out nearly tied on answer quality, but LENS grounded far more of its answers in actual retrieved text - 84.0% versus 70.7%.
That distinction matters more than it sounds. Most retrieval-augmented systems bet everything on an index built before the question exists, which works fine for lookup but breaks down when documents update constantly or when the right level of detail - a table, a paragraph, a whole section - can't be known in advance. LENS makes that tradeoff itself the design choice instead of an afterthought.
It's not a clean win. LENS gives up a few points of raw exact-match accuracy in exchange for evidence it can actually point to, a reasonable bet for anyone who cares whether an answer is grounded rather than merely plausible - but it's a bet, not a breakthrough.