Recall AI

Demo : https://youtu.be/qarlVuubBfM

A minimalist retro-terminal personal journal with semantic search, AI-powered weekly analysis, freeform journal Q&A, and PDF export.

---

What It Does

Journal OS is a local desktop app for writing and reviewing daily journal entries. Every entry you save is automatically embedded into a local vector database (ChromaDB). At the end of each week, an AI agent semantically searches your writing and produces a structured summary. You can also ask your journal any freeform question and get a grounded, cited answer.

Tech Stack

Layer	Technology	Role
UI	Tkinter	Native Python GUI — window, widgets, event loop
Structured storage	PostgreSQL	Persists all journal entries (title, content, date)
DB driver	psycopg2	Python ↔ PostgreSQL connector
Vector store	ChromaDB	Local persistent vector DB for semantic retrieval
Embeddings	OpenAI text-embedding-3-small	Converts journal text to dense vectors
AI orchestration	CrewAI	Agent, Task, Crew; exposes search_journal as a tool
LLM backend	OpenAI GPT-4o	Reasoning and synthesis inside the agent
PDF generation	ReportLab	Renders the styled summary PDF

---

Project Structure

journal_os/
│
├── main.py             ← Entry point. Bootstraps DB, syncs vectors, starts UI.
├── config.py           ← All constants: credentials, model names, ChromaDB path,
│                         colours, fonts. Edit only this file to reconfigure.
│
├── db.py               ← PostgreSQL layer: schema, insert, fetch queries.
│                         Returns entry IDs so the caller can embed immediately.
│
├── vector_store.py     ← ChromaDB wrapper: upsert, semantic query, bulk sync.
│                         Pure retrieval logic — no AI, no UI.
│
├── ai_summary.py       ← RAG-agentic layer: CrewAI agent + search_journal tool.
│                         Exposes generate_summary() and answer_question().
│
├── pdf_export.py       ← ReportLab PDF generation. Takes a string, writes a file.
│
└── ui/
    ├── __init__.py
    ├── widgets.py       ← Stateless widget factories (button, entry, text, etc.).
    ├── chrome.py        ← Persistent app shell: header, live clock, status bar.
    └── screens.py       ← All four screens. Calls db/vector_store/ai_summary —
                           never touches SQL or AI directly.

Dependency graph

main.py
  ├── db.py              (PostgreSQL)
  ├── vector_store.py    (ChromaDB)
  └── ui/
        ├── chrome.py
        └── screens.py
              ├── db.py
              ├── vector_store.py
              ├── ai_summary.py  →  vector_store.py (RAG tool)
              └── pdf_export.py

config.py  ←  imported by everything

---

Setup

1. Prerequisites

• Python 3.11+
• PostgreSQL running locally (default port 5432)
• An OpenAI API key

2. Create the database

CREATE DATABASE journal_db;

3. Install dependencies

python3 -m pip install -r requirements.txt

On Debian/Ubuntu, Tkinter is provided by the system package manager:

sudo apt-get install python3-pip python3-tk

4. Configure credentials

Copy the template to a local .env file and fill in your credentials:

cp .env.example .env

OPENAI_API_KEY=sk-...
POSTGRES_DB=journal_db
POSTGRES_USER=postgres
POSTGRES_PASSWORD=your_pg_password
POSTGRES_HOST=localhost
POSTGRES_PORT=5432

You can also set the same values in your shell:

export OPENAI_API_KEY="sk-..."
export POSTGRES_DB="journal_db"
export POSTGRES_USER="postgres"
export POSTGRES_PASSWORD="your_pg_password"
export POSTGRES_HOST="localhost"
export POSTGRES_PORT="5432"

5. Run

python3 main.py

On first launch, any existing PostgreSQL entries are synced into ChromaDB automatically (idempotent upsert). After that, every new entry is embedded the moment you press Save.

---

How the RAG Agent Works

User presses [ GENERATE SUMMARY ]
        │
        ▼
  App loads the most recent journal entries from PostgreSQL
        │
        ▼
  Agent receives a grounded reflection task with those entries as evidence
        │
        ├─► Optional: search_journal(...) for a targeted theme check
        │         └─► ChromaDB cosine search → relevant older chunks
        │
        └─► GPT-4o synthesises from recent entries first
                  └─► Reflection: Snapshot / Moments / Patterns /
                       Attention / Next Step / Question

If semantic search is unavailable, Generate can still create a reflection from the recent PostgreSQL entries. The Ask feature still requires ChromaDB because it answers open-ended questions through semantic search.

User asks a question
        │
        └─► Tool call: search_journal(...)
        │         └─► ChromaDB cosine search → top-6 relevant chunks
                  └─► GPT-4o answers from retrieved context only

---