XANA OS — PROMETHEUS

Next-Generation Personal Intelligence Platform

"What if your computer actually knew you?"

---

  ██╗  ██╗ █████╗ ███╗   ██╗ █████╗      ██████╗ ███████╗
  ╚██╗██╔╝██╔══██╗████╗  ██║██╔══██╗    ██╔═══██╗██╔════╝
   ╚███╔╝ ███████║██╔██╗ ██║███████║    ██║   ██║███████╗
   ██╔██╗ ██╔══██║██║╚██╗██║██╔══██║    ██║   ██║╚════██║
  ██╔╝ ██╗██║  ██║██║ ╚████║██║  ██║    ╚██████╔╝███████║
  ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═══╝╚═╝  ╚═╝     ╚═════╝ ╚══════╝
                                             P R O M E T H E U S

---

What in the world is this thing?

Glad you asked. XANA OS is a self-hosted intelligence dashboard that lives entirely on your machine. No subscriptions. No cloud. No one watching your queries. Just you, a terminal, and a system that quietly becomes more useful the longer you use it.

It started as a side project. It turned into something that's kind of hard to describe without sounding unhinged. So here's the honest pitch:

XANA OS watches the sky (literally — real aircraft and satellites), listens to the internet, knows your conversation history better than you do, and talks back. It's part command center, part personal AI, part OSINT workstation. It runs on a potato if needed.

What's inside:

• 🌍 Live 3D Globe — real aircraft transponders (OpenSky), vessel positions (AIS), satellite tracks (CelesTrak), all rendered on a pydeck globe in your browser

• 🧠 AI Chat with Memory — conversational AI that actually remembers your past conversations via ChromaDB vector search. Not a gimmick. It retrieves relevant chunks at query time and injects them as context. Your data never moves.
• 🕵️ OSINT Center — world news, known exploited vulnerabilities (CISA KEV), IP/domain recon, crypto markets, weather. All free. No keys needed.

• 👻 PHANTOM Protocol — autonomous multi-source entity investigation. Give it a target. Walk away. Come back to a briefing.

• 🗺️ Neural Map — 3D semantic topology of your entire conversation history. Equal parts useful and unsettling.

All data sources are free and require zero API keys (except optional Windy webcams, which are also free).

---

The Stack

What	How
Web framework	Streamlit
LLM (runs locally)	Ollama — llama3.2 or whatever model you prefer
Vector memory	ChromaDB + all-MiniLM-L6-v2 sentence embeddings
3D globe	pydeck
Charts	Plotly
Graph analysis	NetworkX
Satellite tracking	CelesTrak TLE + Keplerian orbital propagation

No cloud inference. No telemetry. No subscription tier where features disappear.

---

Getting Started

The Fast Path

git clone https://github.com/sriharideveloper/xanaops.git
cd xana-os
./setup.sh        # creates virtualenv, installs deps, checks Ollama
./start.sh        # launches the app

Open http://localhost:8501 and you're in.

What works immediately (no setup): GLOBE, OSINT, all live-feed modules. What needs your data first: ORACLE, ARCHIVE, NEURAL MAP, DOSSIER, CHRONOS — these are the memory-linked modules. See Step 4 to build your brain.

---

Step 1 — Two Prerequisites

• Python 3.10+ — you probably have this
• Ollama — local LLM runtime from ollama.com. Takes two minutes to install.

# Pull a model — the default
ollama pull llama3.2

# Or go smaller (less RAM), or bigger (better answers):
# ollama pull qwen2.5:0.5b   ← potato mode
# ollama pull mistral         ← solid all-rounder
# ollama pull deepseek-r1:8b  ← reasoning tasks

Change the active model in config.py whenever you want.

---

Step 2 — Clone

git clone https://github.com/YOUR_USERNAME/xana-os.git
cd xana-os

---

Step 3 — Install & Launch

./setup.sh        # one-time: creates venv, installs everything
./start.sh        # start XANA OS

Windows or manual setup:

python -m venv xana_env
xana_env\Scripts\activate        # Windows
# source xana_env/bin/activate   # Linux/macOS
pip install -r requirements.txt
streamlit run app.py

First launch downloads the all-MiniLM-L6-v2 embedding model (~90 MB). Get a coffee.

---

Step 4 — Build Your Memory (This Is the Good Part)

This is optional, but it's why XANA OS exists. The memory system ingests your ChatGPT conversation history, embeds it into a vector database, and gives the AI instant semantic search over everything you've ever discussed with a language model.

Think of it as giving your AI a long-term memory it was never supposed to have.

4a — Export Your Conversations

From ChatGPT:

1. Settings → Data Controls → Export Data
2. Download and extract the ZIP
3. Drop all conversations-*.json files into the project root

4b — Parse the Raw JSON

python legacy/parse_chats.py

Reads your conversations-N.json files → produces master_chats.csv. Adjust the file count in the last line of parse_chats.py if you have more than a few.

4c — Pair Messages

python legacy/1_prep_data.py

Takes each user message + AI response and pairs them into semantic blocks → paired_chats.csv. These become your memory units.

4d — Build the Vector Database

python legacy/build-brain.py

Embeds everything using all-MiniLM-L6-v2 and writes it to xana_memory_db/. Grab another coffee. CPU-only is fine — this runs once.

4e — Restart

./start.sh

ORACLE, ARCHIVE, NEURAL MAP, DOSSIER, and CHRONOS are now operational. They will now quietly know things about you that you've forgotten.

---

Step 5 — Windy Webcams (Totally Optional)

Live webcam feeds from anywhere in the world. Free key at api.windy.com.

export WINDY_API_KEY=your_key_here
./start.sh

Or in .streamlit/secrets.toml:

WINDY_API_KEY = "your_key_here"

---

How the Memory System Works

XANA needs no fine-tuning. No model training. No ML degree.

The approach is deceptively simple: instead of teaching the model your data, it retrieves the most relevant chunks from your history at query time and injects them directly into the prompt. The model reasons over context it was handed — not weights it was baked with.

Your question
     │
     ▼
ChromaDB Vector Search  ──►  top N semantically similar memory chunks
     │
     ▼
[System prompt] + [Memory chunks] + [Recent chat history] + [Your question]
     │
     ▼
Local LLM via Ollama  ──►  a grounded, specific, actually-useful answer

Why this is better than you'd expect:

• Works with any Ollama model — swap in config.py, no rebuild needed
• Your data never touches an external server
• No GPU required for training — inference only
• Feed it any ChatGPT export without reformatting

---

Configuration

Edit config.py. That's it. No YAML archaeology.

Setting	Default	What it does
LLM_MODEL	llama3.2	Ollama model name
DB_PATH	./xana_memory_db	Where ChromaDB lives
COLLECTION_NAME	xana_memories	ChromaDB collection
EMBEDDING_MODEL	all-MiniLM-L6-v2	Sentence transformer model
GLOBE_DEFAULT_SAT_GROUPS	stations, starlink, gps, military, weather	Default satellite layers on globe
AGENT_SAFE_COMMANDS	see config.py	Shell commands ORACLE can execute

Switching models is literally one line:

LLM_MODEL = "mistral"   # done

---

Performance Tuning

XANA runs on anything. Here's how to tune it to your hardware without guessing.

The Dials

What	Where	Default	Effect
LLM model	config.py → LLM_MODEL	llama3.2	Speed vs. answer quality
Embedding model	config.py → EMBEDDING_MODEL	all-MiniLM-L6-v2	Search quality vs. build speed
Memory depth	app.py ~line 907 → n=5	5	How many memory chunks retrieved
Context window	modules/database.py ~line 80 → max_chars=4000	4000	Characters of memory fed to LLM
Chat history	app.py ~line 929 → messages[-4:]	4	How many prior turns stay in context

---

🥔 Potato — Old laptop, CPU-only, under 8 GB RAM

Keep it lean. Keep it alive.

config.py:

LLM_MODEL       = "llama3.2:1b"        # or "qwen2.5:0.5b"
EMBEDDING_MODEL = "all-MiniLM-L6-v2"   # already lightweight, keep it

app.py (~line 907):

docs, metas, _ = query_memories(collection, prompt, n=3)

app.py (~line 929):

for msg in st.session_state.messages[-2:]:

modules/database.py (~line 80):

def build_context_string(docs, metas, max_chars=1500):

---

⚡ Mid-Range — Modern CPU, 8–16 GB RAM

You're already here. The defaults are tuned for you. Nothing to change.

LLM_MODEL       = "llama3.2"
EMBEDDING_MODEL = "all-MiniLM-L6-v2"
# n=5 · messages[-4:] · max_chars=4000

---

🔥 Beast — GPU, 16 GB+ VRAM

Push everything up. You earned it.

config.py:

LLM_MODEL       = "llama3.1:8b"           # or mistral, qwen2.5:14b, deepseek-r1:8b
EMBEDDING_MODEL = "all-mpnet-base-v2"      # higher quality, slower DB build (worth it)

app.py (~line 907):

docs, metas, _ = query_memories(collection, prompt, n=10)

app.py (~line 929):

for msg in st.session_state.messages[-8:]:

modules/database.py (~line 80):

def build_context_string(docs, metas, max_chars=8000):

---

Quick Reference

Tier	Model	n=	History	max_chars
🥔 Potato	llama3.2:1b	3	2 turns	1500
⚡ Mid-range	llama3.2	5	4 turns	4000
🔥 Beast	llama3.1:8b+	10	8 turns	8000

Tip: The embedding model only affects DB build time and search quality — not response speed. You can always rebuild with a better embedding model by deleting xana_memory_db/ and re-running build-brain.py.

---

Modules

Module	What it does	Needs DB?
GLOBE	3D live globe — aircraft, vessels, satellites, threat overlays, webcams	No
ORACLE	Memory-linked AI chat with agentic commands	Yes
PHANTOM	Autonomous multi-source investigation protocol	Yes
ARCHIVE	Raw semantic vector search over your memory	Yes
NEURAL MAP	3D semantic topology — visualize your conversation history as a landscape	Yes
DOSSIER	AI-generated intelligence profile compiled from your memory	Yes
OSINT	News, cyber threats, IP/domain recon, crypto prices, weather	No
CHRONOS	Temporal pattern analysis — find out what you used to care about	Yes

---

ORACLE Commands

Type these directly in the chat interface:

Command	What happens
play [song] on youtube	Opens YouTube search
google [query]	Opens Google search
open [app]	Launches an application
weather [city]	Fetches current weather
ip [address]	IP geolocation lookup
lookup [domain]	Domain recon sweep
phantom [target]	Full PHANTOM investigation
recon [target]	Quick recon sweep
exec [command]	Safe shell execution (see config for allowed commands)
status	System diagnostics

---

Data Sources

Everything below is free. Nothing requires an API key.

Source	What it provides
OpenSky Network	Live aircraft transponder positions
Finnish Digitraffic AIS	Real vessel positions (Baltic Sea + global)
CelesTrak	Satellite TLE orbital elements
CARTO	Dark map tiles
Google News RSS	World news headlines
CISA KEV	Known exploited vulnerabilities feed
ip-api.com	IP geolocation
CoinGecko	Cryptocurrency market data
Open-Meteo	Weather forecasts
GDELT	Geopolitical event database

---

Scripts at a Glance

Script	Purpose
setup.sh	One-time setup — creates venv, installs deps, checks Ollama
start.sh	Launch XANA OS

Memory pipeline (legacy/):

Script	Purpose
parse_chats.py	Extracts messages + timestamps from ChatGPT JSON exports
1_prep_data.py	Pairs user/AI messages into semantic blocks
build-brain.py	Embeds everything into ChromaDB

---

Privacy

XANA OS runs entirely on your machine.

The only external network calls are read-only requests to the public data sources listed above. No personal data is transmitted in any of them.

Your conversation history, memory database, and LLM inference all stay on-device. Ollama runs locally. ChromaDB is a folder on your disk. There is no analytics, no telemetry, no phoning home.

What you put into XANA stays in XANA.

---

Screenshots

---

License

MIT — see LICENSE for the legal version of "do whatever you want."

---

ALL SYSTEMS NOMINAL
XANA OS · PROMETHEUS · v∞

---

Built for people who think their tools should be smarter than their cloud subscriptions.