智能新闻编辑室

Installation

git clone https://github.com/alfadur7/llm-wiki-newsroom.git
cd llm-wiki-newsroom

Or click "Use this template" to create your own wiki repo from this scaffold. To start from a clean slate, delete the example documents under raw/ and the example pages under wiki/ (keep the folders and graph/cluster_labels.json), then ingest your own sources with /wiki-ingest.

This project assumes an environment where the AI agent reads and edits files and invokes tools on its own. Support levels by agent:

Claude Code-only features include cascading updates that refresh related existing pages whenever a new document is added, a backlink index across all pages, automatic contradiction tracking, and associative discovery that follows connected concepts. AGENTS.md and GEMINI.md carry only the basic workflow from the original SamurAIGPT project. The Python tools used to build the wiki run locally with no external API keys.

---

Local semantic search (optional)

The python tools/query.py qmd ... subcommand and semantic search in Obsidian and Claude Code are powered by a local search engine called QMD. The embedding and reranking models all run on-device; no data leaves to an external service. If you'll only use graph traversal (the graph subcommand), you can skip this.

1) Install QMD + initial indexing

npm install -g @tobilu/qmd
qmd collection add "$(pwd)/wiki" --name wiki   # register the wiki folder as one collection
qmd embed                                       # generate embeddings (auto-downloads the GGUF model)

The config file is created at ~/.config/qmd/index.yml. This project configures it to exclude from search the root meta files (overview.md, index.md, contradiction.md, etc.) and the auto-generated catalogs (sources/_catalog*.md) — focusing only on source/entity/concept pages with real content to reduce noise.

collections:
  wiki:
    path: /path/to/llm-wiki-newsroom/wiki
    pattern: "**/*.md"
    ignore:
      - "*.md"
      - "sources/_*.md"

2) Use in Claude Code

Place a .mcp.json at the project root registering QMD as an MCP server, and Claude Code automatically invokes semantic search when you run /wiki-query. This file contains local paths, so it's in .gitignore and managed per personal environment.

3) Use in Obsidian

In Obsidian, open this project's wiki/ folder as a vault — it must match the collection path above for search results to link to the correct pages.

It's not in the official community store, so install obsidian-qmd via BRAT (a beta-plugin installer). After installing, apply two customizations for a smoother local-search experience at once with python tools/patch_obsidian_qmd.py:

• Display search-result titles using the frontmatter title: value instead of the chunk heading (## Summary, etc.)
• Configure Hybrid search to automatically skip the heavy LLM stages (reranker, query expansion) — cutting response time from minutes to ten-odd seconds on integrated GPU/CPU setups. When you know the exact term, Keyword mode is fastest at 1–2 seconds (the plugin reloads the model for each search, so going below that requires a separate server that keeps the model resident).

The patch script is idempotent — just re-run it once each time BRAT updates the plugin.

On Windows, after installing Node.js, running the patch script above prepares the Node runtime the plugin needs and also prints the path to put in the plugin settings' Executable Path field. Paste the printed path into that field and search will work.

---

Core workflow

Claude.ai mobile integration

/wiki-export exports the wiki in two layers — RAG (synthesis · directory) synthesizes the answer, and detailed originals are provided by the graph browser (deep links). A Claude.ai project loads attached knowledge wholesale into context rather than retrieving it, so entity/concept bodies (about 70% of the total) and source originals are not put into RAG — doing so would vastly exceed the context limit (~200K tokens). Instead, index.md is a directory holding every entity/concept with a one-line description + deep link, and the synthesis layer (domain overviews · contradiction analyses · analysis reports) fills in the substance of answers. README.md carries the upload budget guide (start from Core ~130K tokens) and the deep-link convention to paste into project instructions. Push these files to GitHub and connect them as Claude.ai Project Knowledge, and you can query the wiki even from a phone — answers built from the synthesis layer, details pointed to via graph links.

So that answers understand the wiki structure and attach links to the graph, paste the entire wiki-export/README.md that export also produces (an instruction document covering file structure, answer rules, and the deep-link convention) once into the custom instructions box of the Claude.ai project. Then, when citing a hub the answer links to that page, and when pinning a specific claim's source it links straight to the original.