Skip to content

Architecture: My Independent AI

Overview

My Independent AI is a modular, privacy-first RAG (Retrieval-Augmented Generation) system. It lets you ask natural language questions over your own personal data — emails, documents, chats — entirely on your own hardware, with no data sent to external AI providers.

Core Principles

  1. Privacy First — PII is scrubbed locally before data is stored. A local SQLite mapping table allows de-anonymization only at read time.
  2. Fully Local by Default — LLMs and embeddings run via Ollama. The vector database is Qdrant running on your machine.
  3. Modular Importers — Each data source (Gmail, WhatsApp, Dropbox, etc.) is an independent plugin following the BaseImporter contract.
  4. Cloud-Optional — GCP (Cloud Run, GCS) is supported for running importers on a schedule, but is not required.

Repository Structure

/
├── apps/
│   ├── admin-dashboard/    # Streamlit UI — chat interface
│   ├── importers/          # Data ingestion plugins
│   ├── orchestrator/       # Ingestion pipeline runner
│   └── qdrant-sidecar/     # GCP VM sidecar for cloud Qdrant
├── libs/
│   ├── embedding/          # Ollama-based embedding client
│   ├── vector-storage/     # Qdrant client wrapper
│   ├── privacy-core/       # PII scrubbing (Presidio) + SQLite mapping
│   ├── content-processing/ # PDF/text extraction helpers
│   └── db-sync/            # SQLite + Litestream utilities
├── infrastructure/
│   └── gcp/                # Terraform for Cloud Run, GCS, VM (optional)
├── scripts/                # Operational helpers (deploy, maintenance)
├── docker-compose.yml      # Local full-stack (Ollama + Qdrant + Dashboard)
└── docs/                   # Architecture, ADRs, contributing guide

Component Diagram

┌─────────────────────────────────────────────────────────┐
│                    User Browser                         │
└───────────────────┬─────────────────────────────────────┘
                    │ http://localhost:8501
┌───────────────────▼─────────────────────────────────────┐
│           Admin Dashboard (Streamlit)                   │
│  • Chat UI with streaming responses                     │
│  • Multi-turn conversation memory                       │
│  • De-anonymization via privacy-core                    │
└──────┬────────────────────────────────┬─────────────────┘
       │ embed query                    │ chat completion
┌──────▼──────┐               ┌─────────▼───────┐
│   Ollama    │               │     Ollama      │
│ nomic-embed │               │   llama3.2      │
│   :11434    │               │    :11434       │
└─────────────┘               └─────────────────┘
       │ vector
┌──────▼──────────┐
│     Qdrant      │
│ personal_data   │
│   collection    │
│    :6333        │
└─────────────────┘
       ▲
       │ upsert embeddings
┌──────┴──────────────────────────────────────────────────┐
│          Importers (run on schedule or manually)        │
│  gmail · whatsapp · synology-nas · dropbox · calendar  │
└─────────────────────────────────────────────────────────┘

Data Flow

  1. Import — An importer fetches data from a source (Gmail API, WhatsApp export, etc.), scrubs PII using privacy-core, and upserts embeddings into Qdrant via libs/embedding + libs/vector-storage.
  2. Query — The dashboard embeds the user's question with nomic-embed-text, retrieves the top-k relevant chunks from Qdrant, then passes those chunks + conversation history to llama3.2 to generate a grounded answer.
  3. De-anonymize — The response is passed through privacy-core's MappingDB to replace anonymized entity IDs (e.g. <PERSON_a1b2c3>) with the original names before display.

Local Stack (Docker Compose)

Service Image Role
ollama ollama/ollama LLM + embedding inference
ollama-init ollama/ollama One-shot model pull (llama3.2, nomic-embed-text)
qdrant qdrant/qdrant Vector database
dashboard Built from apps/admin-dashboard/Dockerfile Web UI

Cloud Deployment (Optional)

For running importers on a schedule without leaving your laptop on, see infrastructure/gcp/. This provisions:

  • Cloud Run Jobs — one per cloud importer (Gmail, Calendar, Dropbox)
  • GCS bucket — intermediate staging area for imported data
  • GCP VM — hosts a persistent Qdrant instance
  • Cloud Scheduler — triggers importers on cron schedules