Xerum compared with a RAG pipeline you build yourself
RAG is a technique. This is the store it should run over.
Chunking, embedding and ranking are a solved fortnight's work, and if that is all you need you should build it. The year of work is underneath: an append-only record, content addressing, a curation loop, permission filtering inside the query, and per-customer isolation.
What building it yourself is genuinely good at
The first version is fast
Embeddings over a document set with a vector store in front of it is a weekend, and it will demo well. That is not a criticism — it is why so many teams start there, correctly.
You control every knob
Chunk size, model, reranking, prompt. When retrieval quality is your differentiator, owning the whole pipeline is the right call and no vendor will tune it better than you.
No procurement, no vendor
It runs on infrastructure you already have, under terms you already agreed to, and nobody has to approve a new supplier.
The same question, in each
What is being retrieved over?
Whatever you pointed it at — usually a drive or a wiki, stale entries included.
A curated record: resources from sources, plus what an owner approved.
Can you answer as of a past date?
Only if you built versioning into the store yourself. Most pipelines index current state.
Yes. Reading at a revision is native, because nothing is ever edited in place.
Where does a citation point?
A chunk id, usually — which is not something you can show a person.
The revision, the resource and the origin URL a human wrote it at.
How are permissions handled?
Commonly post-filtered after ranking, which leaks the existence of what it hid.
Pre-filtered inside the index query, before anything is ranked.
How does anything get corrected?
Somebody edits the source document and you wait for the next re-index.
A proposal, decided by an owner, joining the history as an ordinary change.
What does re-indexing cost?
A full re-embed, typically, because chunks have no stable identity.
Only what changed. Chunks and embeddings are content-addressed and reused.
Keep your pipeline
If you have tuned retrieval for a domain nobody else understands, that work is yours and it is valuable. The question worth asking is what it is running over. Un-curated retrieval inherits whatever is in the swamp, and no amount of reranking fixes a corpus with four versions of the truth in it. Versioning, curation, provenance and permissions are properties of the store — you cannot bolt them onto the retriever afterwards, which is the part most teams discover in month nine.
Skip the year, keep the fortnight
Built today, not on a roadmap. Every property above is already how the store works.