What it does

OpenKB is a CLI tool that ingests documents—PDFs, directories, URLs—and compiles them into a local wiki of summaries, concept pages, and cross-references. The wiki persists as Markdown; queries and chat sessions read from this compiled structure rather than raw documents. It also exports “Skills,” portable expert modules that Claude Code, Cursor, and similar tools can install natively.

The interesting bit

The project explicitly rejects the RAG pattern of rediscovering knowledge per-query. Instead it follows an Andrej Karpathy concept: LLMs do the work once, upfront, building a structured corpus that accumulates and cross-references over time. For long PDFs it uses PageIndex, a vectorless hierarchical tree index, so the LLM reads document summaries rather than choking on full text.

Key highlights

• No vector database; retrieval is reasoning-based via PageIndex tree structures
• Handles both short docs (full-text) and long PDFs ≥20 pages (hierarchical summaries)
• Multi-LLM support through LiteLLM, with models set via provider/model format
• Skill Factory distills wiki subsets into redistributable Anthropic Skills with validation, eval, and rollback
• Chat interface with slash commands (/add, /skill new, /lint) for iterative refinement without leaving the session

Caveats

• Recompiling regenerates summaries and rewrites concept pages; manual edits to the wiki get overwritten
• Requires LLM API keys and ongoing API costs for compilation, query, and chat
• The “no Vector DB” claim is architectural, not a performance benchmark; no latency or accuracy comparisons are provided

Verdict

Researchers, analysts, or teams drowning in PDFs who want a queryable, accumulating memory rather than a search engine should try this. If you need real-time ingestion at scale or guaranteed consistency without LLM hallucination, the compile-everything approach may frustrate.