Open Source Release

ATLAS CE

The world's first open-source deterministic RAG system.
Same query. Same results. Same order. Every time.

What Is ATLAS CE?

Every RAG system in production today has the same flaw: ask the same question twice and you get two different answers. Different documents. Different ranking. Different results. Standard vector search is approximate by design — and approximate means non-deterministic.

That's fine for a chatbot. It's not fine for medical records, financial audits, legal search, or critical infrastructure. You cannot put something you cannot reproduce in front of a regulator.

ATLAS CE solves this with three retrieval lanes fused by locked weights — producing the same results in the same order, every time. We call it LRAG: Lore Retrieval Augmented Generation.

⬡

Semantic Search

Vector similarity for meaning-based retrieval

◈

Keyword Search

BM25 lexical search for keyword precision

▦

Metadata Filter

Exact structured field matching

CE Limitations — Full Transparency

— No proprietary LoreToken compression. CE uses LoreToken-Lite. The semantic compression and GPU memory multipliers of full ATLAS are not here.

— Text and Markdown only. No PDF, DOCX, OCR, or vision pipeline.

— No full Semantic Firewall. Basic adversarial screening is included; the three-stage firewall is not.

— No confidence thresholds or citation enforcement.

The deterministic retrieval architecture — the hard part — is fully here. The rest is up to you.