The hype version of a company knowledge base is a glowing 3D brain floating on screen. The useful version has no picture at all. It is a system that reads everywhere your team already writes, the chat threads, the wiki, the code, and lets anyone ask a plain question and get an answer drawn from all of it. Nick Saraev walked through one that a chip company, Cerebras, runs internally and wrote up in detail, fielding about 15,000 questions a day.
01Put the answer up front
Retrieval augmented generation is just fetching the facts, and the whole game is finding the right facts to put there, and a knowledge base is the machine that finds them. This is the automatic version of what you do by hand when you Brief the model like a brilliant new hire.
02Gather what your team knows
A knowledge base is only as good as what you feed it, so feed it where the knowledge already lives. Slack messages, the wiki of checklists and onboarding pages, the code and its history, and any database of numbers you track.
03Store more than the raw text
Label each item. A chat thread does not go in as a wall of text: a model reads it first and attaches who spoke, when, what they were asking, and how it resolved, etc.
Those labels let the knowledge base weight what it finds. A founder’s message from a minute ago should count for more than a junior’s note from three years back, and a question about last week’s seminar should surface last week’s, not all of them equally. Weighting by recency and source is the difference.
04Match on words and on meaning
When a question comes in, the knowledge base has to find the right saved item, and matching on meaning alone is not enough. Cerebras tried that first and it fell short, so it now runs two kinds of match together.
Exact word matching catches an error string, a host name, or a flag someone pasted in, where the precise characters are the whole point. Let the newer answer win a tie, and each one covers the other’s blind spot.
“how effectively can you retrieve the highest quality ones, the most relevant ones”
Nick Saraev, “Cerebras Just Killed Second Brains” · 22:0505Build it with a coding agent
Coding agents can build the ingestion pipelines, the plumbing that pulls data in, for Slack, email, GitHub, etc.
You supply the understanding of how the data should flow; the agent supplies the wiring.
A knowledge base is the team-sized version of what one agent gets when you Give your agent a memory.
“a system that allows us to store information before the question”
Nick Saraev, “Cerebras Just Killed Second Brains” · 4:37Further reading
- Cerebras · How we built our knowledge base, the primary write-up by Isaac Tai, Daniel Kim, and Mike Gao, with the hybrid search and reranking in full depth
- Nick Saraev · Cerebras Just Killed Second Brains, the walkthrough this guide is drawn from
- Brief the model like a brilliant new hire, the by-hand version of putting the right context in front of the question
- Give your agent a memory, the single-agent cousin of a team knowledge base