IMA Karpathy LLM Wiki Knowledge Base

Problem

Many personal knowledge bases remain piles of fragmented PDFs, notes, and web snippets. At query time, the model may re-search raw documents every turn, producing answers that are hard to audit, hard to link, and blind to contradictions, orphan pages, or missing concepts. ima-wiki-karpathy treats Tencent ima as a durable knowledge store and compiles source material into structured Wiki pages instead of re-running raw-document RAG on every request.

How It Works

The skill follows a Karpathy-style three-layer model: raw/ holds immutable source material, wiki/ holds LLM-generated, cross-linked Markdown pages, and CONFIG.md defines page templates, directory rules, and negative constraints. The main operations are:
- Ingest: reads a PDF, file path, or pasted text, combines it with CONFIG.md and the current index.md, asks the LLM to emit JSON page updates, and writes or updates notes through the ima API.
- Query: retrieves the most relevant Wiki pages, answers with citations such as [source: page title], and explicitly says when the Wiki has no relevant information.
- Lint: scans pages for contradictions, isolated pages, and gaps, then produces a health-check report.
- Convert: extracts PDF text, tables, and metadata with pdfplumber, renders structured Markdown, and optionally continues into Ingest.

Boundaries

This skill fits knowledge bases with a stable topic and long-term accumulation. The quality of CONFIG.md strongly shapes output: the more protocol-like the schema, the more consistent the generated pages. Batch operations should respect API rate limits and include retries. Export Markdown before significant edits as a backup. If the goal is one-off retrieval without maintaining structured pages, the workflow is heavier than necessary.