YouOS ✉️

Your email. Your model. Your style.

YouOS is a local-first AI email copilot that learns from your sent Gmail history and drafts replies that sound like you — not a generic AI. It runs entirely on your Mac. No cloud. No subscriptions. Your data never leaves your machine.

Gmail (sent mail)          Your feedback
       │                        │
       ▼                        ▼
  Ingestion pipeline      Review Queue
  (gog CLI + SQLite)      (10 emails/batch)
       │                        │
       ▼                        ▼
  Reply Pairs DB  ──────► LoRA Fine-tuning
  (FTS5 + BM25)           (Qwen, nightly)
       │                        │
       ▼                        ▼
  Draft Generation ◄──── Autoresearch
  (local Qwen MLX)        (80 iterations/night)
       │
       ▼
  Draft Reply ✅

Privacy: Everything stays local. Your corpus, model, and drafts never leave your Mac.

🌐 Landing page

What it does

• Ingests your sent Gmail history, Google Docs, WhatsApp exports — including organic pairs you sent without YouOS
• Learns your writing style — richer persona: bullet point rate, directness score, sentence length, paragraph style; EWMA-weighted toward recent emails
• Persona re-analysis is incremental (recent 90 days × 3 weight), with full weekly refresh; confidence intervals (p25/p75) shown in prompts
• Per-sender-type personas: different voice, length, greeting, and closing for internal, external client, and personal contacts
• Sender-type style anchors: explicit prompt slot ([STYLE ANCHOR — internal|client|personal]) to stabilize first-draft tone by audience
• Per-account corpus isolation — drafts for work emails draw from work history; personal from personal
• Greets people by first name, closes in your style — greeting and closing injected from persona config per contact type
• Classifies multi-intent (meeting + urgent, etc.), boosts matching exemplars; per-intent reply length calibrated from corpus
• Drafts grounded in score-ranked few-shot exemplars (confidence-annotated, thread-deduplicated); exemplar reply text preserved (600 chars), inbound trimmed (400)
• Exemplar cache by intent+sender-type (TTL + feedback-triggered invalidation) improves consistency and reduces repeated ranking churn
• Prompt token budget enforced — exemplars auto-trimmed if prompt exceeds 2000 tokens
• Confidence thresholds are relative (mean±σ of retrieval scores), not hardcoded
• Subject line + topic-aware retrieval; FTS queries expanded with email vocabulary synonyms
• Same-thread history gets a 2x retrieval boost
• Handles full email threads — paste the whole thread, YouOS focuses on the latest message
• Optional reply instructions — steer a specific draft with explicit guidance even when replying to inbound emails
• Warns you when confidence is low; explain any draft inline via "How was this generated?"
• Local model empty or signature-only output → automatic Claude fallback
• Subject line generated via smart content analysis — skips greeting/filler lines, extracts the actual topic
• Improves from your feedback via LoRA fine-tuning — quality-filtered, deduplicated, curriculum-ordered, DPO preference pairs supported
• Training export deduplicated by inbound similarity (≥0.95 → keep higher-rated pair)
• Auto-scales training hyperparameters; nightly pipeline skips steps when data is insufficient
• Golden eval runs nightly after fine-tuning — composite score tracked in pipeline log
• Autoresearch benchmarks rotate weekly (seeded re-sample) to prevent overfitting to fixed test cases
• Self-optimizes nightly via autoresearch — configurable composite weights, sender-type boosts, intent signals
• Style drift detection: Stats dashboard flags when your writing patterns shift significantly
• Feedback loop closes: high-rating, low-edit pairs surface higher in future retrievals
• Streak tracking — consecutive daily Review Queue sessions tracked; streak shown in queue UI
• Corpus scan button in Stats — bulk-extracts structured facts from your top reply pairs in one click
• Language-filtered retrieval — retrieval matches language of the inbound email; no cross-language bleed
• Sender profiles track reply-time patterns and topics; notes trigger immediate profile rebuild
• Embedding cache for fast repeated retrieval; corpus health at a glance: youos corpus
• Run a golden benchmark anytime (10 curated cases): youos eval --golden
• Runs entirely locally on Apple Silicon

Requirements

• Apple Silicon Mac (M1/M2/M3/M4)
• 8GB+ RAM (16GB recommended)
• Python 3.11+
• gog CLI configured with your Gmail account(s)
• ~5GB free disk space

Quick start

# Clone and install
cd ~/Projects/youos
pip install -e .

# Run the setup wizard
youos setup

# Or run directly
python3 scripts/setup_wizard.py

The setup wizard walks you through:

1. Dependency check
2. Gmail account configuration
3. Email corpus ingestion
4. Writing style analysis
5. Optional initial fine-tuning
6. Server setup

Usage

# Draft a reply
youos draft "Hi, can we schedule a call next week to discuss the proposal?"

# Draft with sender context
youos draft --sender [email protected] "email text here"

# Open the web UI
youos ui

# Check system status
youos status

# View corpus stats
youos stats

# Full corpus health report (pair count, quality scores, top senders)
youos corpus
youos corpus --json   # raw JSON output

# Add a sender note (immediately rebuilds their profile)
youos note [email protected] "prefers bullet points, decision-maker"

# Submit a feedback pair directly from the terminal
youos feedback --inbound "email text" --reply "your reply" --rating 4
youos feedback --inbound "..." --reply "..." --sender "[email protected]" --note "too formal"

# Run nightly pipeline manually (with step-by-step output)
youos improve --verbose

# Check system requirements
youos doctor

# Run golden benchmark evaluation (10 curated cases)
youos eval --golden

# Start the web server
youos serve

# Ingest a WhatsApp export
youos ingest --whatsapp ~/Downloads/WhatsApp-Chat.txt

Facts & Auto-Extraction

YouOS stores contextual facts about your contacts, projects, and preferences to improve draft quality over time. Facts are injected into generation prompts automatically.

Auto-extraction from notes: When you add a sender note (youos note [email protected] "...") or submit feedback with a note, YouOS automatically extracts structured facts using a rule-based extractor with LLM fallback.

How it works:

• Rule-based (primary): finditer over 15+ pattern categories; all matches captured per note
• Negation awareness: Detects preceding negation words (not, don't, never, etc.) and skips false positives
• Confidence scoring: Each pattern carries a base confidence (0.6–0.9); long/noisy captures are downgraded to 0.4
• Fact merging: Duplicate facts (same type + key + text) are deduplicated before upsert
• LLM fallback: When rule-based extraction returns nothing, the Claude CLI is invoked to extract facts from unstructured text

Pattern categories supported:

• Communication preferences: prefers short replies, prefers bullet points, prefers formal tone
• Dislikes / avoidances: hates X, don't like X, never CC X
• Scheduling: meetings on Mon/Wed, available on Fridays, responds within 2 hours, unavailable on X
• Timezone: UTC+5, GMT-8, EST, America/New_York (IANA-style)
• Identity: title/role: X, works at X, based in X, preferred name: X, reports to X
• Sign-offs: signs off with "Best,", use "Cheers" as sign-off, signature: X
• Language: writes in Spanish, speaks French
• Contact metadata: phone: X, billing email: X, always CC X, CC their assistant X
• Relationship tags: decision maker, gatekeeper, VIP client, key account, referred by X
• Project facts: deadline: X, budget: $X, renewal date: X, stakeholder: X

API endpoints:

GET    /api/facts          — list all facts (optional ?type= filter: contact | project | user_pref)
POST   /api/facts          — create or upsert a fact
DELETE /api/facts/{id}     — delete a fact by id

Example fact types:

• contact — key: [email protected], fact: Prefers Tuesday meetings
• project — key: project_alpha, fact: Uses React 18 with TypeScript
• user_pref — key: sign_off, fact: Always close with "Best,"

Facts are stored locally in the SQLite memory table and surfaced via the web UI.

Web UI

The web UI provides:

• Draft Reply: Paste an inbound email (or full thread), generate a draft grounded in your style. A confidence reason banner explains why the draft received its confidence score (e.g. "3 strong exemplars found", "low retrieval — new topic"). See the full exemplar trace via "How was this generated?"
• Review Queue: Emails appear instantly, drafts stream in one by one as they generate. Automated senders filtered by address and content. Configurable batch size (5/10/20) and draft model (claude/local/auto). Keyboard shortcuts: j submit, k skip, e edit, 1-5 rate, ? help.
• History: Past drafts with intent badges, confidence badges, and edit-distance indicators
• Stats Dashboard: Corpus health, model status, pipeline status (with skipped steps), style drift indicator, benchmark trends, edit distance trend chart, per-sender-type accuracy breakdown, and System Health card (corpus size, last ingestion, embedding coverage, adapter status)
• Gmail Bookmarklet: Injects a floating side panel directly into Gmail — auto-detects sender email from the DOM, add an optional instruction, generate a draft, and click "Insert into Gmail" without leaving your inbox. Submit feedback with star rating from the panel

Architecture

app/
  main.py              # FastAPI application
  api/                 # HTTP endpoints
  core/                # Config, embeddings, sender classification
  db/                  # SQLite bootstrap and migrations
  generation/          # Draft generation (local Qwen + Claude fallback)
  ingestion/           # Gmail, Google Docs, WhatsApp importers
  retrieval/           # FTS5 + semantic search
  evaluation/          # Benchmark scoring
  autoresearch/        # Automated config optimization

scripts/               # CLI tools and pipeline scripts
configs/               # Persona, prompts, retrieval settings
templates/             # Web UI (feedback, stats, bookmarklet)

Privacy

All data stays on your machine. No email content is ever sent to a cloud service unless you explicitly use an external LLM for draft generation (configurable). See PRIVACY.md.

Configuration

All settings are in youos_config.yaml, created by the setup wizard:

user:
  name: "Your Name"
  emails: ["[email protected]", "[email protected]"]

model:
  base: "Qwen/Qwen2.5-1.5B-Instruct"
  fallback: "claude"  # or "none" for fully local

autoresearch:
  enabled: true
  schedule: "0 1 * * *"

Running a Personal Instance

You can run multiple independent instances from the same codebase by pointing YOUOS_DATA_DIR at an instance directory. Each instance has its own database, config files, and user data.

Instance directory layout:

instances/myname/
├── youos_config.yaml     # user settings (name, emails, pin, etc.)
├── var/
│   └── youos.db          # SQLite database
├── configs/
│   ├── persona.yaml      # writing style
│   ├── prompts.yaml      # prompt templates
│   └── retrieval.yaml    # retrieval settings
├── data/                 # ingested corpus (raw + feedback)
└── models/adapters/      # fine-tuned LoRA adapter (optional)

Start a named instance:

YOUOS_DATA_DIR=instances/myname uvicorn app.main:app --host 127.0.0.1 --port 8765

When YOUOS_DATA_DIR is set, YouOS derives the canonical DB path as YOUOS_DATA_DIR/var/youos.db. For safety, startup now rejects mismatched DB paths and unsafe paths (for example Trash locations).

Data Safety Commands

# Run integrity checks (required tables + regression warnings)
youos health-check --json

# Create snapshot
youos snapshot-create --tier manual

# List snapshots
youos snapshot-list

# Restore snapshot (with confirmation)
youos snapshot-restore /full/path/to/snapshot.db

# Dry-run restore
youos snapshot-restore /full/path/to/snapshot.db --dry-run

Instance data directories (instances/*/var/, instances/*/data/, instances/*/models/, instances/*/youos_config.yaml) are excluded from git.

License

Open source. See LICENSE for details.