LLM Wiki

Zenii includes a Karpathy-pattern LLM wiki: a persistent, structured knowledge base that an AI agent maintains from raw sources. Unlike RAG (which re-synthesizes from raw docs on every query), the wiki compiles knowledge at ingestion time — the LLM reads a document once, writes structured pages, and maintains cross-references. Future queries draw on pre-built, interlinked knowledge.

Structure

wiki/
  SCHEMA.md      ← operating manual for the LLM agent
  index.md       ← catalog of all pages (LLM-maintained)
  log.md         ← append-only operation history
  sources/       ← drop raw input documents here
  pages/
    concepts/    ← ideas, techniques, frameworks
    entities/    ← people, orgs, projects, products
    topics/      ← subject areas
    comparisons/ ← side-by-side analyses
    queries/     ← saved answers to important questions

Page Types

Each wiki page has a type field that determines how it is classified and where it is stored.

Type	What it covers	Examples
entity	Named, concrete things that exist: people, orgs, products, tools, models, datasets, events	"Andrej Karpathy", "OpenAI", "GPT-4", "PyTorch"
concept	Abstract ideas, techniques, or patterns — not a named thing, but an idea	"attention mechanism", "zero-shot prompting", "chain-of-thought"
topic	Subject area or domain that organizes related pages — broader than a concept	"natural language processing", "model evaluation", "RAG"
comparison	Side-by-side analysis of two or more entities or concepts	"GPT-4 vs Claude 3", "LoRA vs full fine-tuning"
query	A saved answer to a specific question	"What are the main scaling laws?", "When was RLHF introduced?"

How the LLM extracts them

During ingest, the LLM runs a two-pass analysis:

Entity pass — scans the full source for every named person, organization, product, tool, model, dataset, and project. Each gets its own entity page (err on the side of more pages).
Synthesis pass — identifies abstract ideas and techniques (concepts), organizes them under subject domains (topics), and flags any direct comparisons. If the answer to a question is novel synthesis, it can be saved as a query page.

info

Knowledge is compiled once at ingest time, not re-derived on every query. This is the key difference from RAG.

Quick Start: CLI

# Markdown or plain text
zenii wiki ingest notes.md

# Binary documents — auto-converted via MarkItDown
zenii wiki ingest report.pdf
zenii wiki ingest slides.pptx
zenii wiki ingest data.xlsx
zenii wiki ingest spec.docx

# With a specific model
zenii wiki ingest research.pdf --model anthropic:claude-opus-4

All formats use the same command — no flags needed to indicate file type.

Quick Start: Claude Code

Claude Code reads CLAUDE.md which points to wiki/SCHEMA.md, so it already knows how to operate the wiki.

1. Drop a source document

cp ~/Downloads/paper.pdf wiki/sources/

2. Ingest it

ingest wiki/sources/paper.pdf

Claude Code reads the document, creates/updates wiki pages, updates index.md, and appends to log.md.

3. Ask questions

what does the wiki say about transformer attention mechanisms?

4. Run a health check

lint the wiki

Quick Start: Zenii Agent

Zenii's built-in agent has FileReadTool, FileWriteTool, FileListTool, and FileSearchTool built in — no new routes or configuration needed.

Via curl

SESSION=$(curl -s -X POST http://localhost:18981/sessions \
  -H "Authorization: Bearer $ZENII_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"title":"wiki session"}' | jq -r '.id')

curl -X POST http://localhost:18981/sessions/$SESSION/messages \
  -H "Authorization: Bearer $ZENII_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"content":"Read wiki/SCHEMA.md then ingest wiki/sources/my-article.md","role":"user"}'

Via WebSocket (streaming)

const ws = new WebSocket('ws://localhost:18981/ws');
ws.send(JSON.stringify({
  type: "chat",
  session_id: "...",
  content: "Read wiki/SCHEMA.md then ingest wiki/sources/my-article.md"
}));

Ingesting via the Web UI

Open Knowledge Wiki in the sidebar, click Ingest, then drag-and-drop or paste a file. The gateway writes it to wiki/pages/topics/ and returns the new page slug.

Page Title Resolution

When a document has no YAML frontmatter title: field, the title is derived in this order:

Priority	Source	Example
1	Frontmatter `title:`	`title: "My Page"`
2	First `# Heading`	`# My Page` in body
3	Filename	`GitHub Stars.md` → `"GitHub Stars"`
4	Humanized slug	`my-doc` → `"My Doc"`

To control page type and metadata, include a YAML frontmatter block:

---
title: "My Page"
type: concept        # concept | entity | topic | comparison | query
tags: [rust, async]
updated: 2026-04-09
---

Content here.

REST API

Upload a file (binary or text)

POST /wiki/upload
Content-Type: multipart/form-data

Fields:
  file   (required)  Raw file bytes — any format listed above
  model  (optional)  AI model override, e.g. "anthropic:claude-opus-4"

# curl example
curl -X POST http://localhost:18981/wiki/upload \
  -H "Authorization: Bearer $ZENII_TOKEN" \
  -F "file=@/path/to/report.pdf" \
  -F "model=anthropic:claude-opus-4"

Response: same as /wiki/ingest — { primary_slug, pages[], message }.

Ingest text directly

POST /wiki/ingest
Content-Type: application/json

{ "content": "...", "filename": "notes.md", "model": "..." }

Use /wiki/upload for file uploads (binary or text). Use /wiki/ingest when you already have the text content in memory.

Operations

Operation	Trigger	What happens
Ingest	`ingest wiki/sources/X`	LLM reads doc, creates/updates 5-15 pages, updates index, appends to log
Query	Ask any question about the wiki	LLM searches index, reads relevant pages, answers with citations
Lint	`lint the wiki`	LLM scans all pages for contradictions, stale claims, orphans, missing links

Supported Source Formats

Text formats (no conversion needed)

Extension	Notes
`.md`	Best format — ingested directly
`.txt`	Plain text
`.rs`, `.ts`, `.py`, `.go`, etc.	Code files — great for technical wikis

Binary formats (auto-converted via MarkItDown)

Extension(s)	Format
`.pdf`	PDF documents
`.docx`, `.doc`	Word documents
`.pptx`, `.ppt`	PowerPoint presentations
`.xlsx`, `.xls`	Excel spreadsheets
`.html`, `.htm`	Web pages
`.epub`	E-books
`.zip`	Archives (contents extracted)
`.jpg`, `.jpeg`, `.png`, `.gif`, `.bmp`, `.tiff`, `.webp`	Images (requires OpenAI key for descriptions)

Binary files are automatically converted to markdown before ingestion — no extra flags needed.

Setup: MarkItDown

Install the conversion tool once:

pip install markitdown[all]

For image descriptions and OCR, set OPENAI_API_KEY in your environment.

To use a different binary or custom path, add to config.toml:

doc_converter_bin = "markitdown"  # default; override with full path if needed

tip

Obsidian Web Clipper converts web articles to clean markdown with one click.

Example Session

You: ingest wiki/sources/attention-is-all-you-need.md

Agent: Key takeaways:
- Introduces Transformer architecture (encoder-decoder with self-attention)
- Eliminates recurrence entirely — enables parallelization
- Multi-head attention allows attending to multiple positions simultaneously
- Positional encodings substitute for recurrence to track sequence order
- SOTA on WMT 2014 EN-DE and EN-FR translation at time of publication

Creating/updating pages:
- pages/concepts/transformer-architecture.md (new)
- pages/concepts/self-attention.md (new)
- pages/concepts/multi-head-attention.md (new)
- pages/entities/vaswani-et-al-2017.md (new)
- pages/topics/sequence-to-sequence.md (updated)

Updated index.md (5 new entries). Appended to log.md.

You: what does the wiki say about positional encodings?

Agent: From wiki/pages/concepts/transformer-architecture.md:
Positional encodings are fixed sine/cosine vectors added to token embeddings to inject
sequence position information, since the Transformer has no recurrence or convolution.
[source: attention-is-all-you-need.md]

See also: [[self-attention]], [[multi-head-attention]]

Knowledge Graph API

The gateway exposes a machine-readable graph of all wiki pages and their relationships:

GET /wiki/graph
Authorization: Bearer $ZENII_TOKEN

Response: { nodes: [{ id, title, tags }], edges: [{ source, target, label }] } — one node per wiki page, edges from [[wikilinks]] in page content. Use this to power custom graph visualisations or to analyse page connectivity programmatically.

Tips

Commit after each ingest: git add wiki/ && git commit -m "wiki: ingest <source-name>"
Obsidian graph view: open wiki/ in Obsidian to visualize page relationships
Dataview plugin: query frontmatter metadata across pages with SQL-like syntax
Keep sources focused: split large documents into focused chunks for better page granularity

LLM Compatibility

The wiki works with any LLM that has file read/write tools — no embeddings or vector databases required.

Interface	Works	Notes
Claude Code	Yes	Native — reads CLAUDE.md → SCHEMA.md automatically
Zenii agent	Yes	Uses existing file tools, no config needed
Codex CLI	Yes	Point it at `wiki/SCHEMA.md` as the instruction file
Any agent with filesystem tools	Yes	Tell it to read `wiki/SCHEMA.md` first

Structure​

Page Types​

How the LLM extracts them​

Quick Start: CLI​

Quick Start: Claude Code​

Quick Start: Zenii Agent​

Ingesting via the Web UI​

Page Title Resolution​

REST API​

Upload a file (binary or text)​

Ingest text directly​

Operations​

Supported Source Formats​

Text formats (no conversion needed)​

Binary formats (auto-converted via MarkItDown)​

Setup: MarkItDown​

Example Session​

Knowledge Graph API​

Tips​

LLM Compatibility​