Developer Replaces Repeated Context Explanations for AI With an LLM Wiki Workflow

The author of a technical blog on Habr has shown off an LLM Wiki workflow. Instead of re-explaining context to the model with every request, the model gradually adds to and organizes a knowledge base stored in an Obsidian vault. Several articles from Habr itself served as source material, and task processing runs on Codex.

What the workflow is built from

The system combines several tools with three sequential processes, which the author implemented himself on top of an existing Obsidian vault.

• Knowledge store: an Obsidian vault
• Query processing and generation: Codex
• Source material for the knowledge base: several Habr articles
• Three pipeline processes: ingest (loading materials), query (querying the base), lint (quality checks)
• The author calculated ingest costs using Yandex Foundation Models

Splitting the workflow into three separate processes — ingest, query, and lint — lets the author separate responsibilities. Loading new material, querying accumulated knowledge, and checking its quality run as independent steps rather than as one monolithic operation.

How this differs from standard RAG

A classic RAG setup searches the document base for relevant chunks on every request and inserts them into the prompt. The model remembers nothing between sessions, so context has to be rebuilt from scratch each time. In LLM Wiki, by contrast, the model incrementally extends and structures the vault. The result of each query doesn't just answer the user; it also settles into the knowledge base as a separate wiki layer that can be reused in later queries without explaining the context again.

In the article, the author shows the vault structure itself and examples of what this wiki layer looks like in practice, comparing the approach with standard RAG on convenience and cost. As he describes it, the key difference isn't where the model gets its facts, but what happens to an answer once it's produced: in classic RAG the answer is lost after the session, while in LLM Wiki it stays in the base for future queries.

What risks the approach carries

The author separately examines risks that surface quickly in practice with this approach — in particular, the cost of continuously ingesting new material and the need for a dedicated lint process so the knowledge base doesn't degrade or pile up contradictory entries as it grows.

The author calculated ingest costs not in the abstract but on a specific model — Yandex Foundation Models — which suggests an attempt to assess the approach's economics in practice rather than just describe its architecture in theory.

What it means

The LLM Wiki workflow offers an alternative to traditional RAG. Instead of rebuilding context from raw documents every time, the model maintains a living, continuously updated knowledge base. This may reduce the number of tokens spent explaining context, but it requires separate quality control via lint to keep the base reliable.

For anyone using LLMs on large volumes of recurring questions — whether a personal knowledge base in Obsidian or corporate documentation — the approach shows a practical way to cut the cost of repeating the same context in prompts over and over.