How does an agent actually search over a large, distributed set of documents? Classic retrieval-augmented generation chunks documents, embeds the chunks, and pulls the top-k nearest vectors. That works on a demo corpus. At scale — many stores, multi-hop questions, stale indexes — it silently misses. Scroll to see exactly where it breaks, then watch an agentic loop fix it.
A document is split into fixed-size chunks, each chunk is embedded, and each embedding lands as a point in vector space. Drag the sliders: at small chunk sizes with low overlap, a key fact gets split across two chunks — and neither chunk alone answers the question.
Two hundred chunks from four documents, projected in 3D (drag to rotate). Pick a query: the query vector flies in, its top-k nearest neighbors light up cyan — and chunks that a human would call relevant, but that sit far away in embedding space, pulse coral. Those are semantic near-misses. Toggle distributed corpora to see how separate indexes create blind spots.
Question: "What changed in the Q3 refund policy and who approved it?" The answer needs a policy chunk from one store and an approval email from another. Watch both strategies attempt it.
Facts that span a chunk boundary are cut in half. Neither half embeds close enough to the question, so the fact is unretrievable even though it is in the corpus.
k is fixed before you know how much evidence the question needs. Multi-part questions need evidence beyond slot k, which is silently dropped.
Embeddings are computed at ingest time. Documents edited after indexing return outdated chunks with full confidence until a re-index runs.
One-shot similarity cannot follow references: policy points to an approval email points to a ticket. Each hop needs a new, informed query.
Real corpora live in many indexes with different embedding models and permissions. A single nearest-neighbor search cannot rank across incompatible spaces.
An agent plans, issues targeted searches per store, reads results, follows references, re-queries with new terms, and stops when evidence is sufficient — retrieval becomes a loop, not a lookup.
Five quick questions. Instant feedback with explanations.