Raw capture — raw/cerebras-knowledge-base.md
url: https://xcancel.com/cerebras/status/2077822555159945507 title: How We Built Our Knowledge Base author: Cerebras (@cerebras, authors: Isaac, Daniel, Zenghao) date_fetched: 2026-07-25 date_published: 2026-07-16 site: X/Twitter (via XCancel)
Cerebras announcement thread for their technical blog post on building Cerebras Knowledge, an internal RAG system handling 15,000+ employee queries per day. The full technical blog is available at https://www.cerebras.ai/blog/how-we-built-our-knowledge-base
Thread date: July 16, 2026. 2.5M views, 4,577 likes, 447 retweets.
Kevin Simback argues the system is well-designed but surprised Cerebras took a retrieval approach over an ontology-based system. His follow-up thread defines a "shared context layer" / "company brain" as a derived, citable fact layer over all communications and documents — distinct from systems of record. His four-layer architecture: Evidence (raw data) → Facts (entities and claims) → Judgement (reasoning across evidence) → Access (what AI can reason over). Argues retrieval-based systems will not be reliable on questions like "why was this invoice paid when it didn't match the PO?" — they find text but don't solve claims.
Landed on almost exactly the same architecture a month prior: unified Postgres embeddings, hybrid retrieval, reranker, MCP serving LLM-free retrieval primitives. Went further with: append-only canonical store before indexing, formal provenance on every derived artifact, zero LLM calls in the query path, a verified output layer that makes hallucination inexpressible (empty-never-nearest contract, abstention typed as Found|Gap), and governance via the ARSIA Protocol aligned with EU AI Act and GDPR. Full stack on Cloudflare (Workers, R2, D1, Vectorize, Workers AI) for ~$5-10/month.