WikiVault Lightweight File Knowledge Base

Problem

In many agent workflows, docs, web pages, notes, and JSON fragments are scattered across directories. Keyword search can find snippets, but it is weak at composing cross-file answers. Vector databases add embeddings, index services, and operational cost, which can be overkill for a few hundred text files. WikiVault targets this lightweight knowledge-management case by treating the file store as a readable, maintainable wiki rather than a black-box vector index.

How It Works

The skill organizes work around four actions: /wiki-init, /wiki-ingest, /wiki-query, and /wiki-lint. Initialization selects a Node.js or Python runtime and generates wiki/, raw-sources/, and index.md. Ingestion detects duplicates, confirms title, tags, and topic, writes raw files, and compiles the index. Querying has the LLM read the index, locate articles, read the source text, and synthesize an answer. Linting uses scripts for deterministic checks such as missing directories, broken [[wikilink]] references, orphaned pages, and tag quality, then lets the model add semantic review.

Boundaries

It is best for structured or semi-structured text such as md, txt, html, and json, with files ideally under 1MB. Images and PDFs should be extracted to text before import. Binary files and files over 10MB are not suitable. Avoid over-splitting topic values, because cross-topic [[wikilink]] references are more likely to break. If an ingested source references other files, either import those files or explicitly treat the references as examples.