LLM Wiki Agent
Turn source files into a maintained, interlinked local Markdown wiki.
Per-dimension scores and reasoning
Evidence shows: repository provides config files like CLAUDE.md/AGENTS.md, but least privilege principle is not explicit; user confirmation only appears when asking to save query answers as synthesis pages, not covering all operations; data flow transparency is limited, only describing inputs and outputs without internal details; sensitive data handling is not mentioned, but local operation and no API key reduce risk; dependency security is addressed by pinning litellm version in requirements.txt to avoid supply chain attack, but pyproject.toml has broad ranges; external effects include writing files and generating graph.html, but not clearly communicated; rollback relies on git but no dedicated mechanism; source attribution is only via wiki pages, not explicit. Deductions: least privilege, user confirmation, data flow transparency, sensitive data handling, external effects, rollback, source attribution are insufficient or partially missing.
Evidence shows: README and CLAUDE.md etc. are consistent, but no actual tests; dependency availability lists optional dependencies but does not explain handling of installation failures; failure messages are not exemplified. Deductions: self-consistency is acceptable but lacks test evidence, dependency availability not fully explained, failure messages missing.
Evidence shows: clearly targets research, reading, personal knowledge base, business intelligence, competitive analysis scenarios; capability boundaries are defined via supported formats and tools; triggers are explicit (ingest, query, lint, build graph); environment fit covers multiple CLIs and Obsidian integration. Deductions: capability boundaries not explicitly limited, triggers may conflict with other commands, environment fit not covering all platforms.
Evidence shows: information architecture is clear with directory structure; install notes are detailed; naming stability lacks versioning strategy; examples are abundant, FAQ missing; known limitations not explicitly listed; license is MIT; version number exists but no changelog; maintenance responsibility not clear. Deductions: naming stability, known limitations, versioning/changelog, maintenance responsibility are insufficient.
Evidence shows: output is structured wiki pages, directly usable; marginal value lies in automatic knowledge base construction, saving manual effort; cost-benefit is local operation without API key, but optional dependencies. Deductions: output usability not tested, marginal value depends on user scenario, cost-benefit not quantified.
Evidence shows: README claims are not backed by tests; cross-source verification not mentioned; fact vs inference separation not explicit. Deductions: claim traceability insufficient, cross-source corroboration missing, fact-inference separation not explained.
- Dependency ranges are broad; pin versions to avoid supply chain risks.
- User confirmation is incomplete; operations may execute automatically, so be cautious.
- Sensitive data handling is not described; extra assessment needed if processing personal data.
- External effects (e.g., file writes) are not clearly communicated; may modify files unexpectedly.
- Rollback relies on git but no dedicated recovery mechanism.
What does this agent do, and when should you use it?
LLM Wiki Agent is a local knowledge-base workflow for coding agents: place material in raw/, invoke ingest, and let the agent maintain wiki/. It produces source summaries, entity and concept pages, a living overview, an operation log, and saved synthesis pages connected with [[wikilinks]]. The graph/ directory holds graph.json and a browser-openable graph.html; the documented stack is NetworkX, Louvain, Claude, and vis.js. It is documented for Claude Code, Codex, OpenCode, and Gemini CLI through their respective configuration files, with no API key required for the basic agent workflow. It fits people or teams building a durable, versionable Markdown reference instead of re-deriving answers from source documents on every query.
A user runs commands such as ingest raw/papers/my-paper.md, ingest report.pdf, or ingest raw/mixed-folder/, after which the agent reads the supplied material and maintains wiki/index.md, wiki/log.md, wiki/overview.md, and pages under wiki/sources/, wiki/entities/, wiki/concepts/, and wiki/syntheses/. Non-Markdown material is converted through markitdown during ingest; documented formats include PDF, DOCX, PPTX, XLSX, HTML, TXT, CSV, JSON, XML, RST, and EPUB. query: what are the main themes? synthesizes from existing wiki pages; lint checks for orphan pages, broken links, missing entity pages, contradictions, and knowledge gaps; build graph builds the graph from [[wikilinks]]. Graph construction first creates EXTRACTED edges from wikilinks, then has the agent add confidence-scored INFERRED or AMBIGUOUS edges, applies Louvain community detection, and uses SHA256 caching to avoid reprocessing unchanged pages.
- A researcher collecting papers, reports, and articles over several weeks can ingest them under raw/papers/ and accumulate linked pages for methods, organizations, and terminology.
- A reader importing book chapters as Markdown can query changes in a character's motivation, themes, or contradictions in an author's argument.
- A personal knowledge-base user can ingest journals, articles, and podcast notes to gather evidence around concepts such as Sleep, Exercise, and Deep Work.
- A team can ingest meeting transcripts, roadmaps, and customer interviews to identify repeated requests, decisions, and their stated rationale.
- A competitive-intelligence researcher can ingest company announcements and market reports, then save a time-bounded landscape synthesis.
What are this agent's strengths and limitations?
- It treats maintained Markdown pages—not raw retrieval chunks—as the durable unit, accumulating source, entity, concept, and overview pages.
- It distinguishes EXTRACTED wikilink edges from confidence-scored INFERRED and AMBIGUOUS relationships, then emits a self-contained graph.html.
- The documented workflow covers ingestion, synthesis, linting, and graph construction through ingest, query, lint, and build graph.
- It explicitly documents Claude Code, Codex, OpenCode, and Gemini CLI support, plus an Obsidian symlink pattern for browsing the wiki.
- The primary experience requires a coding agent that can read the repository configuration and carry out the documented workflow; it is not presented as a standalone GUI knowledge-base application.
- Automatic multi-format ingestion depends on optional markitdown, while higher-fidelity arXiv, complex-PDF, and batch workflows introduce further Python packages.
- The README describes graph inference as Claude-driven, and standalone scripts in tools/ require ANTHROPIC_API_KEY, creating Anthropic dependence for those advanced or scripted paths.
- Content and generated wiki files live on the local filesystem; the documentation does not describe hosted collaboration, access controls, or a managed service.
How do you install or deploy this agent?
Install Claude Code, Codex, Gemini CLI, or another agent that reads a configuration file, and ensure Git is available. Run:
git clone https://github.com/SamurAIGPT/llm-wiki-agent.git
cd llm-wiki-agentThen start the applicable agent with claude, codex, opencode, or gemini. The documented basic agent workflow needs no API key. To automatically convert non-Markdown files during ingest, install markitdown: pip install markitdown.
How do you use this agent?
Place material in raw/ and enter, for example, ingest raw/papers/my-paper.md in the agent. You can also use ingest report.pdf, ingest slides.pptx notes.docx, or ingest raw/mixed-folder/ for mixed files or a directory. Use query: what are the main themes? to synthesize from the wiki, lint to inspect links and knowledge gaps, and build graph to create graph.html. Claude Code additionally exposes /wiki-ingest, /wiki-query, /wiki-lint, and /wiki-graph; the other documented agents use equivalent natural-language triggers.
How does this agent compare with similar options?
The README contrasts it with RAG: RAG re-derives knowledge at query time and uses raw chunks as retrieval units, while LLM Wiki Agent compiles structured wiki pages, pre-builds cross-references, and flags contradictions during ingest. That makes it a better documented fit for a persistent, browsable knowledge asset; the source does not establish performance or operational trade-offs against dedicated RAG systems for on-demand document retrieval.