Should I build my own RAG pipeline?

Build it when retrieval is your product and you have people to run ingestion quality, evaluation and drift. Buy it when retrieval supports your product and what you need is verifiable answers next quarter rather than a pipeline to maintain.

How is Cortex different from an orchestration framework?

A framework gives you the wiring and leaves ingestion quality, reranking, citation resolution, the PDF viewer, hosting and data posture to you. Cortex is the running service behind two endpoints, with claim-level citations and residency guarantees included.

What is the hidden cost of DIY RAG?

Not the first demo — the second year. Parser regressions on new document types, chunking changes that invalidate evaluations, reranker tuning, citation mapping, PDF highlighting, and the operational load of a vector store.

Can I keep my own model?

Yes. Cortex supports model choice, and the rag endpoint returns ranked evidence you can feed to whichever model you already run.

Comparison

Build it yourself, use a framework, or use the Cortex SDK

Three legitimate routes to retrieval-augmented answers. They differ less in whether they work and more in who carries the provenance and the maintenance.

Short answer: a hand-built pipeline gives maximum control and the highest ongoing cost; an orchestration framework gives wiring but leaves ingestion quality, citation resolution and hosting to you; the Uthereal Cortex SDK gives cited answers and highlighted PDF sources behind two endpoints, with EU/Swiss residency and zero retention.

Create your agent Read the SDK guide

Option 1 — Hand-built pipeline

You run parsing, chunking, embeddings, a vector store, a reranker, prompt assembly, citation mapping and a PDF viewer, plus the evaluation harness that tells you when a change made answers worse.

Best when retrieval quality is your competitive product
Full control over every stage and every model
Cost concentrates after launch: parser regressions, drift, reranker tuning, infrastructure
Citations are the hardest part and usually the part that gets skipped

Option 2 — Orchestration framework

Frameworks give you composable building blocks and a fast first demo. They do not give you ingestion quality, stable passage identifiers, claim-level citation resolution, a highlighting viewer, or a data-residency posture you can show a procurement team.

Fast prototype, familiar abstractions, large ecosystem
You still choose and operate the vector store and reranker
Provenance remains your responsibility
Upgrades across the stack are yours to absorb

Option 3 — Uthereal Cortex SDK

Cortex is the running pipeline. You call ask for a cited answer or rag for ranked evidence; the SDK renders claim-level citations and opens the cited PDF page highlighted.

No vector database, chunking pipeline, reranker or citation UI to build
Claim-level citations resolved from retrieval, not generated as text
Retrieval-only output when you want to keep your own model
EU/Swiss residency, zero data retention, no training on your content
Your AI coding tool can wire it in from published instruction files

The whole integration

const id = await cortex.createConversation();
for await (const update of cortex.ask(id, { message })) render(update);
// citations already resolved — <CitedAnswer /> opens the highlighted page

How to choose

Answers must be auditable by a regulator or a professional reader → provenance-first service
Retrieval quality is the product you sell → build, and staff it properly
You need a working cited feature this quarter → Cortex SDK
You already run a good model and only lack evidence → Cortex rag endpoint

Frequently asked questions

Should I build my own RAG pipeline?: Build it when retrieval is your product and you have people to run ingestion quality, evaluation and drift. Buy it when retrieval supports your product and what you need is verifiable answers next quarter rather than a pipeline to maintain.
How is Cortex different from an orchestration framework?: A framework gives you the wiring and leaves ingestion quality, reranking, citation resolution, the PDF viewer, hosting and data posture to you. Cortex is the running service behind two endpoints, with claim-level citations and residency guarantees included.
What is the hidden cost of DIY RAG?: Not the first demo — the second year. Parser regressions on new document types, chunking changes that invalidate evaluations, reranker tuning, citation mapping, PDF highlighting, and the operational load of a vector store.
Can I keep my own model?: Yes. Cortex supports model choice, and the rag endpoint returns ranked evidence you can feed to whichever model you already run.

RAG API with sources

Criteria and endpoints.

How to add citations

The architecture behind cited answers.

Uthereal Cortex SDK

Full developer guide.

Lovable setup

One prompt to integrate.