Developer replaced repeated context explanation for neural network with LLM Wiki workflow

A technical blog author on Habr demonstrated an LLM Wiki workflow: instead of re-explaining context to a neural network with each request, the model gradually populates and organizes a knowledge base in Obsidian storage. Several articles from Habr itself were used as sources for populating the base, and task processing was built on Codex.

What the Workflow Is Built From

The system relies on a combination of several tools and three sequential processes that the author implemented independently on top of an existing vault in Obsidian.

• Knowledge storage — vault in Obsidian
• Request processing and generation — through Codex
• Sources for populating the base — several articles from Habr
• Three pipeline processes: ingest (loading materials), query (querying the base), lint (quality checking)
• The author calculated the cost of ingest on the Yandex Foundation Models model

Building the workflow around exactly three separate processes — ingest, query, and lint — allows the author to divide responsibility: loading new materials, accessing already accumulated knowledge, and checking quality happen as independent steps, not as a single monolithic operation.

How This Differs From Regular RAG

The classic RAG approach, with each request, newly searches for relevant fragments in a document base and substitutes them in the prompt — the model remembers nothing between sessions, and context has to be reassembled from scratch each time. In LLM Wiki, the model, on the other hand, incrementally supplements and structures the vault: the result of each query not only answers the user, but also settles in the knowledge base as a separate wiki layer that can be reused in subsequent queries without re-explaining the context.

The author shows in the article the vault structure itself and examples of what this wiki layer looks like in practice, comparing the resulting approach with regular RAG in terms of convenience and cost. According to his description, the key difference is not where the model gets the facts from, but what happens with the answer after it is received: in classic RAG the answer is lost after the session, while in LLM Wiki it remains part of the base for subsequent requests.

What Risks Does the Approach Have

The author separately analyzes the risks that quickly manifest in practice with such an approach — in particular, the question of the cost of constant ingest of new materials and the need for a separate lint process so that the knowledge base does not degrade and does not accumulate contradictory entries as it grows.

The author calculated ingest costs not abstractly, but on a specific model — Yandex Foundation Models — which speaks to an attempt to evaluate the economics of the approach in practice, rather than just describing its architecture in theory.

What This Means

The LLM Wiki workflow offers an alternative to traditional RAG: instead of reassembling context from raw documents each time, the model maintains a live, constantly updated knowledge base — this can reduce the number of tokens spent explaining context, but requires separate quality control through lint to keep the base reliable.

For those working with LLM on large volumes of repetitive questions — whether a personal knowledge base in Obsidian or corporate documentation — such an approach shows a practical way to reduce the cost of constantly repeating the same context in prompts.