LLM Wiki Monorepo

<p align="center"> <strong>Agent-native knowledge compiler.</strong><br> AI agents turn raw documents into persistent, cross-linked Markdown wikis.<br> No database. No API lock-in. One <code>git clone</code>. </p> <p align="center"> <a href="#what-is-this">What</a> • <a href="#features">Features</a> • <a href="#quick-start">Quick Start</a> • <a href="#architecture">Architecture</a> • <a href="#packages">Packages</a> • <a href="#templates-20-domains">Templates</a> • <a href="#documentation">Docs</a> • <a href="#credits">Credits</a> • <a href="#license">License</a> </p> <p align="center"> <a href="https://github.com/JeanBaissari/llm-wiki-monorepo/actions/workflows/ci.yml"><img src="https://github.com/JeanBaissari/llm-wiki-monorepo/actions/workflows/ci.yml/badge.svg" alt="CI"></a> <a href="https://pypi.org/project/baissarienterprises-llm-wiki/"><img src="https://img.shields.io/pypi/v/baissarienterprises-llm-wiki.svg" alt="PyPI"></a> <a href="https://python.org"><img src="https://img.shields.io/badge/python-3.10+-blue.svg" alt="Python 3.10+"></a> <a href="https://nodejs.org"><img src="https://img.shields.io/badge/node-18+-green.svg" alt="Node 18+"></a> <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-yellow.svg" alt="License MIT"></a> <a href="https://deepwiki.com/JeanBaissari/llm-wiki-monorepo"><img src="https://deepwiki.com/badge.svg" alt="Ask DeepWiki"></a> </p>

---

pip install baissarienterprises-llm-wiki

No local models required. LLM Wiki is agent-native: the intelligence comes from the agent you already run (Codex, Hermes, Claude, OpenCode). The base install is lexical/keyword-only and performs no model downloads. Optional extras for local computation remain opt-in and are not part of the supported default story.

What's New in v0.6.5 — "CI Green"

Embedding degrades instead of crashing. If the optional [semantic] model cannot load at runtime (offline runner, missing cache), llm-wiki embed now reports the degradation and leaves keyword search untouched instead of raising IndexError.

Provenance scan respects content. Wiki pages and working directories are no longer scanned for ported-code markers, so ordinary prose can't fail the provenance gate.

What's New in v0.6.4 — "Lean & Certified"

Local-model surface removed. The unused @sentropic/graphify hard dependency is gone (with its 239-package / ~606 MiB closure: @ai-sdk/*, Ollama provider, tree-sitters, DB drivers), along with the --code-analysis/merged/export-graph code paths and the web code-graph route (ADR-0035). The knowledge graph is fully offline and wikilink-derived. Optional Python extras remain opt-in and are no longer promoted as a recommended profile.

Provenance closed. graph-engine/src/relevance.ts and insights.ts were independently reimplemented as clean-room MIT code with behavior parity verified against the previous implementation; the provenance ledger now reflects this.

Packaging fixed. install.sh now installs the Python package (the llm-wiki console script actually exists afterwards), and wheels ship all 20 domain templates plus py.typed. scaffold resolves package-internal templates first, so pip-installed wikis are no longer silently generic.

Integrity hardened. Lock acquisition is race-free (O_EXCL + owner token + rename-based stale break), page writes are atomic, backups are temp+rename with sha256 verification and staged restore, malformed tuning.toml yields a clean config error, hybrid search honors --set bm25.*, and no-provider ask fails with an actionable message instead of a silent null.

Release gates are real. The version manifest is green, docs_truth_check scans the live docs, the certifier's gates exit nonzero on failure, the zero-test guard covers all workspaces, a slow-tests lane runs, coverage is floored at 70%, and publishing is gated on certification.

What's New in v0.6.2

Install.sh completeness. install.sh now builds all packages in dependency order: graph-bridge → graph-engine → mcp-server → audit-shared → web-viewer → obsidian-audit. Previously graph-bridge was missing, causing graph-engine imports to fail.

Template scaffolding fix. scaffold.py now handles both array and dict extra-dirs.json formats — 7 templates (business, commodities, decompilers, developer-tools, medicine, personal-growth, reading) previously created wrong directories.

Developer hygiene. .hermes/ added to .gitignore to prevent accidental commits of local skill config.

What's New in v0.6.0 — "Epistemic & Surface"

One-command setup. llm-wiki setup <root> [--title] scaffolds/validates a wiki and registers the MCP server with every detected client — claude, codex, opencode, hermes — idempotently, with --dry-run/--uninstall and a health + tools/list smoke test (LWM_035).

Ask this wiki. llm-wiki ask <root> "<question>" answers questions grounded in the wiki — hybrid retrieval over pages + community summaries with a summary-aware rerank, exactly one structured LLM call on the agent-native $0.00 default, a --no-llm deterministic offline mode, and a faithfulness contract (answer entities ⊆ cited pages). MCP llm_wiki_ask included (LWM_033).

Contradictions + evidence confidence. llm-wiki contradictions <root> detect|list|apply|unapply extracts typed claims, detects contradictions (unit-normalized numerics, polarity, exclusive categories) suggest-only, and computes evidence-grounded confidence (high/medium/low + evidence_score) into the confidence/contested/contradictions frontmatter fields — author-overridable via confidence_source: evidence|author (LWM_034).

Demo wiki. llm-wiki demo <dest> materializes a committed, lint-clean "Redis Internals" playground (8 pages) from the installed package or repo, in one command (LWM_036).

Optional local-compute extras (legacy, not required): [semantic], [leiden], [entity-resolution], and [ner] remain available for advanced local computation but are deliberately outside the supported default story — the base install and the agent-native $0.00 path cover every core operation (LWM_037).

Web-viewer derived overlay + Sigma.js + exports. The quarantined derived layer renders as an off-by-default dashed overlay (byte-identical when off), with a Sigma.js WebGL view and JSON Canvas / JSON-LD exports — web-viewer-only diff, no backend change (LWM_038).

Standing gold-set curation. scripts/curate_gold_set.py + a release-certify gate_search_goldset_fresh gate make per-minor gold-set growth a procedure, not a decision (LWM_039 §A).

What's New in v0.5.0 — "Graph Precision"

Entity resolution. llm-wiki entities resolve|list|unmerge collapses variant surface forms ("GPT-4" / "GPT 4" / "gpt-4") into one canonical id through a reversible canonical↔alias table (JSONL source of truth + additive SQLite alias tables). Every merge is reversible; no page prose is ever rewritten. ER-F1 on a committed gold set gates it (ADR-0024).

Leiden community detection. Optional [leiden] sidecar (graspologic) with hierarchical levels — guaranteed-connected communities, NMI/modularity verification against Louvain, TS still only consumes (no TS Leiden, ADR-0025).

Typed + directed + bitemporal edges. One additive edge-schema evolution (relType/directed/validFrom/validTo/observedAt) — the undirected default output is byte-identical (ADR-0026).

Derived edges, quarantined. The graph discovers similarity + co-occurrence edges into a separate layer, excluded from all analytics by default and included only when the NMI+modularity gate passes (fail-closed, ADR-0027).

Hierarchical community summaries. llm-wiki summarize-communities writes first-class community-summary pages per community level + a global summary, with faithfulness filtering and stale-page cleanup (LWM_030).

Tuning constants → config. Every precision constant (relevance weights + type-affinity matrix, insights thresholds + signal scores, community resolution/seed, RRF k, BM25 k1/b, claim penalties — 52 settable keys) lives in one canonical TuningConfig surfaced by llm-wiki tuning with CLI > env > file > default precedence, emitted to the graph-engine via --tuning-json (ADR-0028).

Hybrid search is the default. Search fuses BM25 + semantic KNN via RRF by default, degrades to keyword byte-identically without the [semantic] extra, and keeps --keyword as the escape hatch — certified by a committed search gold set + gate (ADR-0020).

What is this?

LLM Wiki is a production-grade knowledge engine that turns raw documents into a living, cross-linked Markdown wiki. Instead of re-retrieving documents on every query (RAG), the system incrementally builds and maintains a persistent knowledge base. Sources are compiled once, kept current, and compound over time.

It's everything needed to run a self-building wiki in one monorepo: a Python CLI package, an MCP server, a knowledge graph engine with community detection, a Chrome web clipper, a web viewer, an Obsidian plugin, and 20 domain templates — wired through an agent skill that works with any LLM.

---

Five ways to run this

One repo, five entry surfaces — all reading and writing the same wiki/ directory. Pick the one that fits your setup.

1. CLI (Python package)

pip install baissarienterprises-llm-wiki
llm-wiki search ~/my-wiki "attention"

Scripts, CI, and humans get the full command surface — run llm-wiki <command> --help to explore. See the CLI reference.

2. MCP server (any MCP client)

npx llm-wiki-mcp --wiki ~/my-wiki
claude mcp add llm-wiki -- npx llm-wiki-mcp --wiki ~/my-wiki

Programmatic wiki access for any MCP client (Claude, Codex, Cursor, opencode) via 15 stdio tools. Register it with claude mcp add, the opencode opencode.json mcp form, or llm-wiki setup (one-command wiring, v0.6.0). Requires a built mcp-server: cd mcp-server && npm run build (or bash install.sh). See the MCP tools reference.

3. Hermes skill (in-conversation agent workflow)

ln -sf /path/to/llm-wiki-monorepo/skill ~/.hermes/skills/research/llm-wiki

Loads the ten-operation skill for Claude/Hermes sessions — agent-native, no API keys needed. See skill/SKILL.md.

4. Cron / automation

0 3 * * * llm-wiki ingest ~/wikis/my-project raw/articles/my-source.md --llm opencode

Schedule maintenance like ingest, lint, or backup. portalocker advisory locks make concurrent agent runs safe — multiple agents or CI jobs can operate on the same wiki without corrupting pages. See the concurrency reference and the quickstart.

5. MCP server launcher (stdio)

llm-wiki serve ~/my-wiki

Relay the MCP server over stdio (with --build to compile it on demand and --projects for multi-wiki mode). MCP clients normally launch npx llm-wiki-mcp --wiki <root> directly — see section 2 above. The optional web preview (mermaid, KaTeX, audit feedback) is the separate web-viewer server: cd web-viewer && npm start -- --wiki ~/my-wiki. It is local-only by default — see the security boundary.

---

Features

• Two-Step Ingest — LLM analyzes sources first, then generates structured wiki pages. SHA256 caching skips unchanged files. Streaming progress, multi-provider support (OpenAI, Anthropic, DeepSeek), and agent-native mode that needs zero API keys.

• Agent-Native Provider — route ingest through Hermes, Claude Code, or Codex directly. No external API keys required when running inside an AI agent session.

• Structured Output — Pydantic-typed parsing via instructor. No regex guesswork. Retry wit