Most RAG systems chop documents into chunks and hope the right ones
surface at query time. Nexus does the hard work up front:
it distills your sources into curated artifacts so the
information an answer needs is already organized and easy to find — then
answers questions with Search-as-Code.
The idea: distill, don't just chunk
Raw chunks are noisy: the same fact is scattered across pages, counts
require reading everything, and relationships are implicit. Curation
reorganizes the corpus into artifacts — condensed,
typed units of knowledge — so retrieval gets a head start.
Pre-computed, typed knowledge
Curation produces artifacts by kind — summaries, topics, entities,
events, per-doc and per-page units. Each is small, focused, and
describes itself, so the right one is easy to retrieve.
Structure for exact answers
Structured artifact types become queryable tables, so "how many"
and "list every…" questions get exact answers instead of a guess
from prose.
Relationships made explicit
Edge types connect artifacts (who mentions what, what happened
when), turning a flat pile of text into a graph you can walk.
How answers happen: Search-as-Code
Curated artifacts are only half the story. At query time the agent
doesn't just stuff chunks into a prompt — it writes Python against a retrieval SDK to gather exactly the evidence it needs, then
answers from that.
The model orchestrates retrieval
One turn can read artifacts off disk, run semantic and keyword
search, walk the relationship graph, and query structured tables —
composing several operations instead of a single top-k lookup.
Only evidence enters the prompt
Raw tool output stays in the code sandbox; just the distilled
evidence the agent selected reaches the model. That keeps context
small and answers grounded — every claim traces to a source.
It can iterate
If the first pass isn't enough, the agent refines its code and
searches again — so hard questions get more work, not a worse answer.
Why it matters
Distilling artifacts up front plus Search-as-Code at query time means
higher-quality answers on large, messy corpora: exact counts, real
relationships, and citations you can trust — without dumping everything
into the model's context.