Developer preview
Shape RAG context automatically
from live demand.
Pelorus Query watches repeated user and agent traffic, creates focused Evidence Extracts from your source chunks, and clusters related intents into reusable context packages. Retrieval adapts over time while staying grounded in your corpus.
THE PROBLEM
Chunking is doing two jobs.
Most RAG pipelines start by splitting documents into chunks, embedding those chunks, and matching incoming queries against the vector store. That works, but it forces one representation to serve two different purposes: match live traffic accurately and carry enough context to answer well.
Smaller chunks can match precisely but lose surrounding meaning. Larger chunks preserve context but dilute the semantic signal. Parent chunks, hierarchical retrieval, agentic chunking, reranking, and embedding tuning can improve the tradeoff, but they do not remove it. The retrieval surface is still mostly defined before real demand arrives.
WHAT IF?
Context could be shaped by live demand?
What if retrieval was not limited to matching live queries against document chunks chosen before traffic arrived?
What if live demand could shape a new context layer?
With self-contained, source-grounded information packets designed around
the intents users and agents actually express so that matching could move
from query-to-chunk toward intent-to-intent, while the returned context
stays true to the corpus.
Could that change the operating loop entirely?
So that you could see where demand is forming, where clusters overlap,
where the corpus has no answer, and where high-value context deserves
human curation, without rewriting and reprocessing the original documents
every time.
COMING SOON
Pelorus Query
Demand-shaped context engineering for RAG.
Pelorus Query watches live RAG traffic, turns recurring intents into source-grounded Evidence Extracts, maps emerging clusters and gaps, and lets operators curate the context layer without constantly rewriting, rechunking, and re-embedding the source corpus.
Automated Evidence Extracts
Recurring demand produces self-contained context packages that carry the source-faithful context needed to answer, rather than relying only on the chunk boundaries chosen at ingest.
Intent and cluster matching
Future queries match against learned intents and cluster-level extracts, not only raw document chunks. When clusters overlap, Pelorus can return multiple relevant extracts instead of forcing a brittle single bucket.
Demand, coverage, and gaps
Visualize traffic as it forms, trigger automatic creation of targeted extracts, and see exactly where demand spikes, overlaps, or exposes gaps in the corpus.
Human curation without reingest
Refine high-value extracts, add clarifying context, or fill known gaps in the retrieval layer without editing and reprocessing the original documents every time.
Status
Open-source developer preview.
Pelorus Query is early software for builders exploring retrieval observability, demand-driven curation, and source-grounded evidence workflows.