RAG vs. tacit knowledge

RAG retrieves relevant passages from what your organization wrote and feeds them to the model; tacit knowledge is what your organization never wrote, so RAG cannot retrieve it. The two are not rivals. They do not even meet: one is a pipeline over the written record, the other is everything outside that record.

The comparison is practical, not academic. RAG genuinely fixes the explicit-knowledge problem — a model that ignores your written policies, prices, and procedures. It does nothing for the tacit-knowledge problem: the judgment, workarounds, and unwritten rules your company actually runs on. Teams buy RAG expecting the second fix and get the first.

It retrieves. Before the model answers, a retriever searches your corpus — wiki pages, tickets, contracts, code — for passages relevant to the question. Each document is split into chunks, each chunk becomes an embedding vector, and the question's vector is matched against them; the top few chunks get pasted into the prompt. The model then answers from those passages instead of from its training data alone.

That is the whole trick, and it is a good one. The model never "learns" your documents; it reads a handful of them, per question, at answer time. Which means everything depends on one condition: the passage that answers the question has to exist in the corpus at all.

Grounding answers on documents that exist. On questions your corpus actually answers, RAG is a real improvement: the model quotes your refund policy instead of inventing one, cites the source so you can check it, and picks up yesterday's policy change the moment the index does — no retraining. Access control can ride along, so people only get answers drawn from documents they are allowed to read.

So, no strawman: if your problem is that a general model ignores what your organization wrote down, RAG is the right tool and you should use it. This page is about what happens at the edge of the written record.

Because there is nothing to retrieve. Retrieval searches text, and tacit knowledge has no text. The workaround everyone uses, who to actually ask, why the discount rule exists — nobody typed any of it into a system the indexer can crawl. No document, no chunk; no chunk, no vector; no vector, no match.

The failure is silent, which is what makes it dangerous. A retriever always returns its top k passages — it has no way to report that the real answer is missing from the corpus. So the model receives the closest written thing, a generic policy or a stale doc, and composes a fluent answer from it. Nothing in the pipeline errors. The answer is simply wrong about how you work.

Confident, generic, slightly wrong. Ask the assistant how deals get approved and it recites the process doc — the one from two reorgs ago, or a plausible generic version stitched from whatever the retriever found. It does not know finance signs off anything over 50k, because that rule lives in nobody's document. It answers like a smart new hire on day two: fluent about everything written, oblivious to how things actually run.

The cost is not one wrong answer; it is trust. After the third slightly-wrong answer about their own company, people stop asking the assistant — and quietly go back to interrupting the one person who knows.

No — both scale the same input, and the gap is what never entered that input. A million-token context window lets you paste the entire drive into the prompt; the entire drive still does not contain the workaround. Fine-tuning adjusts weights on training text; where no text exists, there is nothing to train on.

This is the useful diagnostic. Retrieval, long context, fine-tuning, agents that browse your drive — every technique in the stack shares one dependency: source text. They all fail on tacit knowledge at the same place, the input, for the same reason. Which also tells you where the real fix has to happen.

Getting the knowledge into words before retrieval runs — elicitation first, then indexing. The pipeline is not broken; the corpus is incomplete. Once someone writes "we never discount for this client, because of the 2023 renegotiation," RAG handles that sentence as well as any policy doc: chunked, embedded, retrieved, cited.

In practice: interview the people who hold the knowledge and write down what they say. Capture decisions with their why at the moment they are made, not in a wiki sprint two years later. And treat every correction of the assistant — "no, we actually do X, because Y" — as the missing document it is, and file it into the corpus. Unglamorous, human work. It is also the one step no model can do for you, because the source is not text. It is people.

Dark context names the gap RAG structurally can't fill: what an organization knows but never wrote down — the situated context a general AI model cannot see. A retrieval pipeline makes the boundary of the written record unusually visible, and dark context is everything on the far side of that boundary.

The tacit gap is one shape of dark context, not the whole of it. Dark context also covers knowledge that was written and then went dark — the doc that drifted out of date, the thread fenced off in a leaver's mailbox. But the never-written layer is the deepest, and surfacing it — putting it into words — is what makes it retrievable at all. We — Ola Möller and Andriy Zhukov, at Dark Context — coined the term by analogy to dark matter: the unseen mass that holds the visible structure together. Your retriever reaches the visible part.

Can RAG retrieve tacit knowledge?

No. RAG searches text — wikis, tickets, contracts — and tacit knowledge has no text. It was never written down, so there is no chunk to embed, no vector to match, and nothing for the retriever to return. RAG grounds answers on the written record and stops exactly where the record stops.

What is RAG actually good at?

Grounding answers on documents that exist. On questions your corpus answers, RAG cuts hallucination, cites sources you can check, and picks up changes by re-indexing instead of retraining. If a general model is ignoring your written policies, RAG is the right fix and no strawman.

Why does my RAG assistant give generic answers about my company?

Because a retriever always returns its nearest written passages, and for questions about how you actually operate there often are none. It never reports not-found; it returns the closest match — a generic policy, a stale doc — and the model composes a fluent answer from it. The result is confident, generic, and slightly wrong.

Does a bigger context window or fine-tuning fix the tacit-knowledge gap?

No. A long context window lets you paste more of the written record into the prompt; fine-tuning trains weights on the written record. Both scale the same input, and the gap is what never entered that input. Where no source text exists, there is nothing to load and nothing to train on.

How does the RAG tacit-knowledge gap relate to dark context?

Dark context names what RAG structurally can't fill: what an organization knows but never wrote down — the situated context a general AI model cannot see. Surfacing it, putting it into words, is what makes it retrievable at all; once written, ordinary retrieval handles it. The term was coined by Ola Möller and Andriy Zhukov at Dark Context.