Neural Composer

approved

by Oscar Campo

Local Graph RAG powered by LightRAG. Chat with your notes using deep knowledge graph connections. - This plugin has not been manually reviewed by Obsidian staff.

★ 162 stars↓ 12,169 downloadsUpdated 18d agoMIT
Install in ObsidianView on GitHub

Neural Composer

Graph-based AI chat for your Obsidian vault.

Hero Banner

Release License: MIT Obsidian

TL;DR

Chat with your vault using a Knowledge Graph, not just keyword search. Neural Composer runs a local LightRAG server, builds a graph of your notes, and lets you ask questions that trace connections across your entire vault.

  • 🔍 Finds relationships, not just matching words
  • ⚡ Manages the LightRAG server for you — no terminal juggling
  • 🔒 100% local when used with Ollama — your data never leaves your machine

Requirements: Python 3.10+ · pip install "lightrag-hku[api]" · Obsidian 1.7.2+

Features

⚡ Automated ServerStarts and stops the LightRAG Python process automatically. No terminal needed.
🧠 Graph + Vector SearchCombines entity-relationship traversal with semantic vector search for deep, contextual answers.
📂 Vault SyncSet a watched folder — notes are re-indexed on save. Status dots in the file explorer show each note's graph state: 🟢 processed · 🟡 processing · 🔴 failed · 🔵 removed.
📊 Knowledge Graph ViewExplore your graph visually in 2D or 3D. Overview mode renders all nodes; Explore mode does a BFS walk from any entity.
🌐 Remote ServerConnect to a LightRAG instance on a NAS, VPS, or Docker container.
🤖 MCP ToolsExpose your graph to any MCP-compatible client (Claude Desktop, etc.).
🔍 Source TransparencyEvery answer includes citations [1] linked to the exact notes and text chunks that were used.
🔒 Local & PrivateUse Ollama for a fully offline setup, or any hosted provider you prefer.
Complete list of supported file 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

Why Graph RAG?

Standard vector search finds similar text. Graph RAG finds connected ideas.

Standard Vector SearchNeural Composer (Graph RAG)
How it worksFinds chunks that match your query semanticallyTraverses relationships between entities in your notes
Best for"What is X?""How does X influence Y across my research?"
Context qualityOften fragmentedHolistic — sees the whole picture
Multi-hop reasoning✗✓

Getting Started

📖 Full documentation on the Wiki

1. Install the LightRAG backend

# Recommended: use a virtual environment
python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install "lightrag-hku[api]"

Then find the path to the installed executable (you'll need it in Step 3):

which lightrag-server        # macOS / Linux
where.exe lightrag-server    # Windows

2. Install the Plugin

Search for "Neural Composer" in Settings → Community Plugins → Browse and enable it.

3. Connect & Configure

Open Settings → Neural Composer. The panel has a sidebar with seven tabs:

  1. Providers — add your API keys (OpenAI, Anthropic, Gemini, Groq, Ollama, etc.)
  2. Models — select your chat, apply, and embedding models
  3. Graph & Vault — set the lightrag-server path, choose a data directory, and optionally configure a Watched Folder for auto-sync
  4. Toggle Auto-start on, then click Restart Server

A green dot in the status bar confirms the server is running. Right-click any folder in your vault to ingest notes and start chatting.

🛠️ Use Cases
  • Researchers — synthesize arguments across hundreds of papers, surface consensus and contradictions that keyword search misses.
  • Writers & Game Masters — track relationships between characters and lore; keep your world internally consistent without digging through folders.
  • Journalers — connect entries from months ago to today, spotting patterns that aren't visible day-to-day.
  • Project Managers — visualize dependencies between project notes that otherwise look like separate tasks.
🧩 Advanced Options
FeatureWhere to configure
Watched FolderSettings → Graph & Vault → Watched folder
Remote ServerSettings → Graph & Vault → Use remote server
Custom OntologySettings → Graph & Vault → Ontology section — teach the graph domain-specific entity types (e.g. "Experiment", "Theorem")
RerankingSettings → Graph & Vault → Reranking — Jina AI, Cohere, or a custom local endpoint
MCP ServersSettings → Tools (MCP)
Graph VisualizationSettings → Graph & Vault → Graph rendering engine — 2D (fast) or 3D (immersive)
Performance TuningSettings → Advanced — chunk size, overlap, async workers
Custom .env overridesSettings → Advanced — raw .env editor with full LightRAG configuration access
🔒 Privacy & Security

Neural Composer is designed with privacy as a core principle.

What leaves your machine

DestinationWhenWhy
Your AI provider (OpenAI, Anthropic, Gemini, Groq, etc.)Every chat message or ingestionTo generate responses and embeddings. Only notes you explicitly ingest or attach are sent.
Your local LightRAG server (localhost)Every query and ingestionThe plugin talks to a Python process on your own machine. No data leaves.
Your remote LightRAG serverOnly if you configure a remote URLOff by default. Opt-in only.

Using Ollama + local LightRAG = zero data leaves your machine.

What never happens

  • The plugin does not send telemetry, analytics, or crash reports.
  • The plugin does not contact github.com or any external domain at runtime. Links in the UI are navigation-only — never fetched programmatically.
  • API keys are stored only in Obsidian's own data.json in your local vault.

System-level access disclosures

Why the Obsidian scanner flags certain capabilities
CapabilityReason
fs (filesystem)Writes the LightRAG .env config file to your chosen work directory, which may be outside the vault.
child_process (shell)Starts and stops the local LightRAG Python server. The command is always the exact path you configure — no user input is interpolated into shell arguments.
Vault enumerationLists file paths for ingestion and the search index. File content is only read when you explicitly ingest a file.
ClipboardInherited from the Lexical rich-text editor in the chat input. Standard paste operations only.
atob/btoa (Base64)Used by bundled deps: @modelcontextprotocol/sdk decodes JWT tokens for MCP OAuth; sigma/three-forcegraph encode WebGL shader data. No sensitive data is encoded this way.
new FunctionUsed by two bundled libraries: ngraph.forcelayout (3D physics) and ajv (JSON schema validation via MCP SDK). Neither executes user-provided code.
📋 Changelog

v1.4.0 — 2026-05-27

  • Mobile support (iOS / Android) — plugin loads on Obsidian mobile and chats against a remote LightRAG server over HTTP. lightRagUseRemote is forced on, local-server management settings are hidden, and the bundle ships an events polyfill plus a require shim so node-only deps don't abort module evaluation on a non-Electron webview.
  • Graph view on mobile — the "desktop-only" notice is gone; the view renders via the same /graphs HTTP endpoints. A right-anchored sidebar slides in/out via a new toolbar button and an x next to the "Node manager" title. Node sizes shrunk for narrow viewports. Newer LightRAG versions (≥1.4) now use the file_path property sent on each node, so the local kv_store_*.json reads aren't needed for citation filenames on either platform.
  • Fix: AbstractJsonRepository.ensureDirectory() was fire-and-forget — on Android adapter.list() raced ahead of mkdir and crashed the template list. Every public method now awaits a shared directory-ready promise.

v1.3.1 — 2026-05-25

  • Fix: removed all !important CSS declarations — replaced with higher-specificity selectors to comply with the Obsidian plugin linter.

v1.3.0 — 2026-05-24

  • Settings UI redesign — new sidebar navigation with 7 tabs: Providers, Models, Chat, Graph & Vault, Tools (MCP), Advanced, Help.
  • Document status tracking — colored dots in the file explorer for each note (🟢 processed, 🟡 processing, 🔴 failed, 🔵 removed). Watched folder shows an aggregate status dot.
  • LightRAG version detection — the server version is displayed as a badge in Settings → Graph & Vault.
  • Watched folder sync — notes are automatically re-indexed on save with a 5-second debounce.
  • "Remove from graph" action — right-click context menu lets you remove individual notes from the graph without deleting the file.
  • Tooltip improvements — status bar tooltip correctly distinguishes local vs. remote server offline state.

v1.2.3 — 2026-05-22

  • Fix: correct LightRAG provider config for OpenRouter and Ollama (LLM_BINDING_HOST was missing, causing 401 errors).
  • Fix: expose active embedding model selector in settings UI.
  • Add: "Reprocess failed documents" button in Graph & Vault settings.
  • Fix: stop server now correctly kills orphaned processes on macOS/Linux via port lookup.

v1.2.1 — 2026-05-20

  • Knowledge Graph Visualization — 2D and 3D interactive graph view inside Obsidian.
  • Overview mode (all nodes) and Explore mode (BFS from a selected entity).
  • Real relevance scores for cited references (citation-frequency formula).
  • Improved "Context used" panel — shows scores, snippets, and click-to-open for .md files.
  • Fix: single-click on isolated nodes now auto-explores and shows full entity details.

v1.2.0 — 2026-05-17

  • Initial public release on the Obsidian Community Plugin marketplace.
  • Local LightRAG server management (auto-start, restart, stop).
  • Right-click folder ingestion with multi-format support.
  • Chat with graph RAG, hybrid query modes, Jina/Cohere reranking.
  • Custom ontology (entity types) and .env editor.

Built on the shoulders of giants:

For plugin developers

Search results and similarity scores are powered by semantic analysis of your plugin's README. If your plugin isn't appearing for searches you'd expect, try updating your README to clearly describe your plugin's purpose, features, and use cases.