Building a Personal Wiki Without Claude Code: A MapReduce-Based Pipeline

Inspired by Andrej Karpathy's idea of LLM-generated wikis, the author (a Habr user) explores alternatives to Claude Code for building a personal wiki. Three strategies were tested: sequential extraction with context memory, batch extraction with a separate integration agent, and a MapReduce approach. The first failed because weak models lost elements when the count exceeded ten. The second failed due to poor integration that merged unrelated items. The third, MapReduce, succeeded: each source is annotated and key terms extracted, then terms are inverted to map each term to its sources. An LLM judge filters terms by importance; for important terms, a wiki element is created with summary, properties, and links to related elements. The pipeline uses gpt-5-mini from a proxy provider and gemma4:26b from local Ollama. gpt-5-mini tends to overestimate significance and sometimes adds terms on its own, but a linker catches this; gemma4:26b is stricter but struggles with structured output, requiring splits in code. Costs: Claude Code costs $3-4 per 100KB of source text, gpt-5-mini costs $1.1-1.3, while gemma4:26b only costs electricity but is slow (12 tokens/sec, taking a full day for the pipeline). The implementation, Wiki-assembler, is a Python pipeline with five or six simple LangChain agents and two hardcoded blocks, avoiding ReAct due to Gemma's infinite reasoning. The pipeline steps include annotation and categorization, filtering key terms via an LLM judge, building wiki elements, and linking them.