Neural Composer

Deep, graph-based search for your Obsidian Vault.

👋 Hello, Obsidian Community!

We built Neural Composer because we love Obsidian, but we often felt limited by standard search tools.

Have you ever searched for a topic in your vault and gotten a list of notes that contain the word, but miss the context? Or tried to ask an AI plugin a complex question, only for it to fail because it couldn't "see" the connections between your files?

We wanted a way to talk to our notes that felt like talking to someone who actually remembers them.

That's why we integrated LightRAG (Graph-based Retrieval) into Obsidian. Unlike standard plugins that just look for matching text chunks, Neural Composer builds a Knowledge Graph of your ideas, helping you find relationships you might have forgotten.

---

🤔 Why use Graph RAG?

Standard AI search (Vector RAG) is great for finding similar text. But Graph RAG is better for finding connected ideas.

Feature	Standard Vector Search	Neural Composer (Graph)
How it searches	Finds matching keywords/concepts	Follows relationships between entities
Best for	Simple questions ("What is X?")	Complex questions ("How does X influence Y?")
Context	Often fragmented	Holistic and interconnected

---

🛠️ How it helps (Use Cases)

We designed this to fit into different workflows. Here is how it might help you:

• For Researchers: If you have hundreds of papers, you can ask it to synthesize arguments across multiple authors, finding consensus or contradictions that a simple search would miss.
• For Writers & DMs: If you are building a world or a story, the graph tracks the relationships between characters and lore, helping you maintain consistency without digging through folders.
• For Daily Journalers: It connects entries from months ago to today, helping you spot patterns in your life or work that aren't obvious day-to-day.
• For Project Managers: It helps visualize dependencies between different project notes that might otherwise look like separate tasks.

---

Features

We wanted the experience to be as smooth as possible:

• ⚡ Automated Server: No need to fiddle with terminals. The plugin handles the background Python server for you (starts and stops automatically).
• Hybrid Search: You don't have to choose. It combines Vector search with Graph traversal for the best results.
• Easy Ingestion: Right-click any folder to add your notes to the graph. It supports PDFs, DOCX, and more...
  Complete list of supported formats

  md, txt, docx, pdf, pptx, xlsx, rtf, odt, epub, html, htm, xml, json, yaml, yml, csv, tex, log, conf, ini, properties, sql, bat, sh, c, cpp, py, java, js, ts, swift, go, rb, php, css, scss, less
• 🔍 Transparency: The chat shows you exactly which files and text segments were used to generate the answer (with citations like [1]), so you can always verify the source.
• 🔒 Local & Private: You can use local models (like Ollama) for a completely offline experience, or connect to Gemini/OpenAI if you prefer.

---

Getting Started

Full documentation (in construction) on wiki

This plugin requires a small backend setup (Python) to run the LightRAG engine.

1. One-time Setup

1. Ensure you have Python 3.10+ installed.
2. Install the engine via terminal:

   pip install "lightrag-hku[api]"
   

   (We recommend using a virtual environment).

2. Install the Plugin

• Recommended: install via BRAT
• Manual Installation: Download main.js, manifest.json, and styles.css from the Releases page and place them in your .obsidian/plugins/neural-composer folder. Enable it in Obsidian.

3. Connect & Go

Go to Settings > Neural Composer.

1. Enter your API Keys (Gemini/OpenAI/Ollama).
2. In the Neural Backend section, paste the path to your lightrag-server executable and choose a folder for your data.
3. Turn on Auto-start and click "Restart Server".

You are ready! You can now right-click a folder to ingest your notes and start chatting with your vault.

---

🧩 Advanced Options

For those who like to tinker, we added some power features:

• Custom Ontology: Teach the graph the specific categories of your field (e.g., "Experiment", "Theorem") instead of generic ones.
• Reranking: Connect to Jina AI or a local reranker for higher precision results.

---

🔒 Privacy & Security

Neural Composer is designed with privacy as a core principle. Here is an exact account of every network call the plugin makes and why.

What leaves your machine

Destination	When	Why
Your AI provider (OpenAI, Anthropic, Gemini, Groq, etc.)	Every chat message or ingestion	To generate responses and embeddings. Only the notes you explicitly ingest or attach are sent.
Your local LightRAG server (localhost)	Every query and ingestion	The plugin talks to the LightRAG Python process running on your own machine. No data leaves.
Your custom / remote LightRAG server	Only if you configure a remote URL	If you opt in to a remote server, queries go to that URL. This is off by default.

If you use local models (Ollama) and a local LightRAG server, no data ever leaves your machine.

What never happens

• The plugin does not send telemetry, analytics, or crash reports to any server.
• The plugin does not contact github.com or any other domain at runtime. GitHub URLs visible in the settings UI are navigation links only — they are never fetched programmatically.
• The plugin does not store your API keys anywhere other than Obsidian's own data.json in your local vault.

System-level access (Obsidian scorecard disclosures)

The Obsidian automated scanner flags several capabilities. Here is the plain-English explanation for each:

Disclosure	Why it exists
Direct filesystem access (fs)	Required to write the LightRAG .env configuration file to your chosen work directory, which may be outside the vault.
Shell execution (child_process)	Required to start and stop the local LightRAG Python server. The command is the exact path you configure in settings — no user input is ever interpolated into shell arguments.
Vault enumeration	Required to list your notes for ingestion and for the RAG search index. Only metadata (file paths) is read; content is read only when you explicitly ingest a file.
Clipboard access	Inherited from the Lexical rich-text editor used in the chat input. Allows standard paste operations.
Base64 calls (atob/btoa)	Used by bundled dependencies: @modelcontextprotocol/sdk decodes JWT tokens for MCP OAuth, and sigma/three-forcegraph encode WebGL shader data. No API keys or sensitive data are encoded this way.
Dynamic code execution (new Function)	Used by two bundled libraries: ngraph.forcelayout generates optimized N-dimensional physics code for the 3D graph, and ajv (via the MCP SDK) compiles JSON schema validators. Neither is used to execute user-provided code.

---

This project is a labor of love, built upon the shoulders of giants:

• Forked from the excellent Smart Composer by glowingjade.
• Powered by the LightRAG library.
• Developed by Oscar Campo & Cora (AI).

We hope this helps you connect the dots in your own second brain. Happy composing!