RAG re-discovers per query; llmwiki compiles once

What it does

llmwiki ingests raw documents and runs a two-phase LLM pipeline to extract concepts and generate typed pages—concept, entity, comparison, overview—with paragraph- and claim-level citations pinned to exact source line ranges. The result is a browsable, interlinked markdown wiki with a local web viewer, hybrid semantic/BM25/graph retrieval, and an eval harness that scores health, citation precision, and regression deltas. It also exposes an MCP server so Claude, Cursor, and other agents can pull budgeted, citation-aware context packs directly from the compiled artifact.

The interesting bit

The project treats RAG as a bug, not a feature: instead of re-discovering relationships from chunks at query time, it front-loads the work into a compile step that merges duplicate concepts, resolves [[wikilinks]], and bakes the structure into persistent markdown. Saved queries become new wiki pages, so the knowledge base actually gets smarter the more you use it.

Key highlights

• Two-phase compile extracts concepts from all sources before generating any pages, eliminating order-dependence and merging shared concepts into single canonical entries.
• Incremental, content-hash-aware pipelines: only changed sources trigger LLM calls, and embeddings update surgically rather than rebuilding the whole index.
• Citation traceability down to line ranges (^[source.md:42-58]), with a built-in linter to validate every marker and an eval mode that gates CI on precision thresholds.
• Provider-portable LLM backend: Anthropic, OpenAI-compatible local servers (llama.cpp, vLLM), Ollama, and GitHub Copilot.
• MCP server integration serves the compiled wiki to any MCP-compatible agent, including a get_context_pack tool that returns budgeted evidence with full provenance.

Caveats

• The GitHub Copilot provider lacks an embeddings endpoint, so semantic search falls back to full-index selection; switch to OpenAI or Ollama if you need vector retrieval.
• Per-concept prompt budgets default to ~50k tokens and truncate aggressively when sources pile up, printing a warning to stderr so you know which concept hit the budget.

Verdict

Worth a look if you maintain a growing corpus of notes, papers, or design docs and want a structured artifact that outlives the current chat session. Skip it if you only need ad-hoc search over rapidly changing or one-off data where traditional RAG is cheaper and sufficient.