16-4 Building an LLM Wiki 📚

This feature grows scattered material into a knowledge wiki the agent manages on its own. A movie's lore, a recipe book, a paper archive, a work wiki — anything works.

← Home: ObsiKO 101 · Parent: 16 AI Agent

Getting started

(Optional) Open a fresh vault, and tell the agent "build me an LLM wiki."
The agent will ask whether you want a quick start or a deep start. - Quick start: just give it a topic and language, and it sets up a basic skeleton in seconds. - Deep start: it asks more — who you are and why you're collecting this (your reuse axis), the wiki's sub-categories, the kinds of material, how often it updates, and even the agent's personality (persona) — for a sturdier setup.
Once you've answered, it builds the folders, system files, and a getting-started guide (Read-me-first.md) all in one go.

📱 On mobile too (a lightweight edition)

Even without a desktop, you can start a lightweight LLM wiki from the chat AI on your phone or tablet.

Say "build me an LLM wiki" in chat and it creates the essential skeleton — the rulebook (AGENTS.md), index.md, and the 00. Inbox / 10. Raw Sources / 20. Wiki folders. That's already a real wiki: collect, ask, and lint all follow the rules.
Day-to-day use works on your phone as well: the chat recognizes a wiki vault on its own and files material by the rules. "Organize this", questions, "check this over" — all fine.
Requirements: the chat AI has an API key, and the note-editing tools toggle is ON in options (≡). → 16-8 Chat-only Mode (Mobile & Web)
For the full edition (the Soul.md persona interview, /collect-style command files, the wiki-map.canvas structure map), continue with the desktop agent — it's the same vault, so the skeleton you built on your phone carries straight over.

The resulting structure

System files: CLAUDE.md (entry point/manifest) · Soul.md (the agent's identity) · AGENTS.md (working rules) · user.md (about you) · memory.md (things you've agreed on together) · index.md (dashboard index) · log.md (work log)
Folder hierarchy:
00. Inbox/ — an inbox where you drop new material.
10. Raw Sources/ — kept exactly as-is (never edited). Organized by kind, e.g. 11. Articles/12. Papers/13. Books….
20. Wiki/ — the wiki pages the agent curates and maintains. Organized by category, e.g. 21. Concepts/22. Entities/23. Guides/24. Maps….
30. Queries/ — where answers to your questions get filed back as pages.
80. References/ — storage for images and attachments.
A structure mind map, wiki-map.canvas — a mind map of the whole wiki at a glance (center = wiki name, branches = categories). You can drag it around and edit it yourself!

💡 Subcategory names adapt to your topic — a worldbuilding wiki might use 21. Characters/22. Factions/23. Places/24. Timeline instead.

📥 File formats you can drop in Inbox (the agent can read them)

These are the file extensions the agent can read and organize when you put them in 00. Inbox/ or drag them into the chat.

Kind	Extensions	Notes
Text/Markdown	`.md` `.markdown` `.txt`	Read directly, no tools needed
PDF	`.pdf`	Text-based PDFs are extracted as-is; scanned (image) PDFs get OCR (needs `poppler`, Claude only)
Images	`.png` `.jpg` `.jpeg` `.gif` `.webp` `.bmp`	Viewed, transcribed, and described (Claude only)
Office/documents/e-books	`.docx` `.pptx` `.odt` `.epub` `.rtf` `.html`(`.htm`) `.tex` `.rst` `.org` `.wiki` etc.	Dozens of formats supported via Pandoc (needs `pandoc`)
Audio/video (transcription)	`.mp3` `.wav` `.m4a` `.aac` `.flac` `.ogg` `.opus` `.wma` · `.mp4` `.mov` `.mkv` `.webm` `.avi` `.m4v`	Local `whisper` (+`ffmpeg`) or cloud transcription with an OpenAI/Gemini API key
Data (tables)	`.csv` `.tsv` `.json`	Converted into Markdown tables

💡 poppler, pandoc, and whisper are free external tools — they're only needed when you ask for that specific format, and if they're not installed you'll get a one-line install tip. (Text, images, and data tables work right away with no extra install! And transcription works without whisper too — with an OpenAI/Gemini API key it goes through the cloud.) For video URLs, YouTube, or RSS, just give the address instead of a file.

Running it (4 core actions)

Just drop material into 00. Inbox/ and talk to it. It works both in natural language and via slash commands. (Note: /collect, /ask, /lint, /status below are not the built-in assistant commands — they work through rules (system files) installed into that vault when you build an LLM wiki — so calling them in natural language is more reliable. Newly built wikis install the commands for desktop and mobile chat, so /collect shows up on your phone too.)

What you want	Natural language	Command
Organize/incorporate material	"organize this"	`/collect` (ingest)
Ask the wiki a question	just ask	`/ask` (query)
Check and fix	"check this over"	`/lint`
See status	"show me the status"	`/status`

Collect: keeps Inbox material as-is in 10. Raw Sources/, then creates or updates wiki pages linked together with wiki links, and updates index.md and log.md. One topic = one page — it first looks for an existing page (aliases included) and updates that, adding new spellings as aliases; re-adding the same source doesn't create duplicates. Processed Inbox files are marked "done" with a one-line note so they aren't processed twice.
Ask: answers based on the wiki pages, citing sources, and saves valuable answers to 30. Queries/.
Lint: fixes orphaned pages, broken links, and index mismatches; suspected duplicates and contradictions (flagged with > [!warning]) are reported as candidates for you to decide — never merged on its own.
Status: summarizes counts per tier and recent collection activity from the actual files.

Making it your own

Open Soul.md (personality), AGENTS.md (rules), user.md (about you), and memory.md (shared memory) and edit them directly to change how the agent behaves. The wiki is effectively the agent's "brain"! (Wikis built earlier may use a hidden .memory.md instead — it works the same.)
Start by reading the Read-me-first.md file created in the vault.
(Optional) In the panel's options (≡) button → the 'git' category, turn on "Auto-commit to git after agent work" to stack up changes as snapshots — connect a remote (GitHub) too for automatic backups! → 23 Settings & Themes

💡 Tip — Use it like NotebookLM, a "my-material" assistant

Just like Google NotebookLM, you can use it as an assistant that answers only from material you've collected and cites its sources. (The tools and semantic search above effectively play the role of RAG here.) If setting up a full LLM wiki feels like too much, you can start light like this instead.

Gather material first: put the PDFs, documents, images, audio, and notes you'll draw on into one folder (e.g. 00. Inbox/ or Material/). Any extension from the "file formats you can drop in Inbox" table above works.
Answer only from your material, with sources cited: just add one line to your question — - "Answer only based on the attached/this folder's material, and cite the source (filename, passage). If it's not in the material, say so honestly (no making things up)."
The same kinds of outputs NotebookLM gives you: - "Make a core-summary briefing from this material" · "Turn it into a study guide (concepts, likely questions)" · "Make 10 FAQs" · "Make a timeline" · "Make a comparison table of the differences and similarities between these two sources"
Finding by meaning (semantic search): ask something like "find everything similar to this topic," and it'll use a local index to find passages that are close in meaning even if the wording differs (and answer based on those results). → See 16 AI Agent for how to turn it on

ObsiKO's edge: both your material and the answers stay as your own .md files, so they're truly yours — and if you want to grow it further, you can promote it straight into an LLM wiki and start running /ask (source-cited Q&A) and /lint. In other words, it's not a "use once and done" notebook, but a knowledge base that keeps growing.

← Previous: Writing & Composition · Home · Next: Zettelkasten & Ontology →