Async Researcher

Your vault's memory and research engine, running locally.

Lumen turns an Obsidian vault from a passive store of notes into an active research collaborator. It runs entirely on local hardware — vault contents never leave the machine.

What it does

Proactive surfacing — as you open or edit a note, the plugin pushes related notes, forgotten threads, and connections you wouldn't have made manually into a side panel.

Deep research on demand — ask a complex question and the system retrieves relevant vault content, synthesises a long-form report, and verifies every citation against a real ^id block in your notes. No hallucinated references. Progress streams live as the local model works.

Cross-session memory — the system remembers what you've researched before and uses that context to focus future research plans.

Architecture

Three processes at runtime:

Obsidian plugin  ↔  FastAPI orchestrator (:8765)  ↔  Ollama (:11434)

• obsidian-plugin/ — TypeScript plugin (esbuild). The only HTTP client. Handles surfacing side panel (related notes), deep research modal with live SSE progress, and report note writing.
• apps/orchestrator/ — Python FastAPI backend. Thin routes; logic lives in flows/. Runs the always-on file watcher in-process.
• packages/ — Python libraries: ingestion, retrieval, citation, inference, memory, web.
• eval/ — eval harness + datasets for validating retrieval and citation quality.

Quick start

# One-time setup
cp config.toml.example config.toml   # edit [vault].path
make install                          # uv sync + pnpm install

# Pull models (adjust to taste)
ollama pull qwen2.5:14b-instruct
ollama pull qwen2.5:7b-instruct
ollama pull nomic-embed-text

# Initial vault ingestion
make ingest

# Start the orchestrator (includes file watcher)
make dev

# Build and symlink the plugin into your vault
make plugin-build
ln -s "$(pwd)/obsidian-plugin" "<vault>/.obsidian/plugins/lumen"
# Then enable it in Obsidian → Settings → Community plugins

Commands

make install       # install all dependencies (Python + Node)
make dev           # start FastAPI orchestrator + file watcher
make ingest        # one-shot bulk vault ingestion
make plugin-build  # build obsidian-plugin/main.js
make plugin-dev    # esbuild watch mode
make test          # run Python tests + plugin vitest
make lint          # ruff + tsc --noEmit
make format        # ruff format + autofix

Tech stack

• Python 3.12 — uv workspaces, FastAPI, LightRAG, DuckDB, LangGraph, mem0ai
• TypeScript — Obsidian API, esbuild
• Ollama — local inference via OpenAI-compatible API
• pnpm — Node package manager

Planning docs

All architectural reasoning lives in docs/:

Source specs (ArchitectureSpecification.md, LocalAgenticRAGArch.md) at the repo root are the north star — not a literal build list.