DeskChat — A Private Desktop AI Assistant for Your Documents

DeskChat is a fully local, privacy-first desktop chatbot that can understand and answer questions about your own files.

It runs entirely on your machine and combines:

• A normal conversational chatbot
• Retrieval-Augmented Generation (RAG) over your documents
• A multi-agent reasoning pipeline (Answerer → Critic → Refiner)
• Optional free web search (DuckDuckGo)
• Built-in document tools (summaries, TOC, citation extraction)
• A clean web UI built with Streamlit
• Optional secure remote access via ngrok

Nothing is sent to any cloud AI service. All models run locally using Ollama.

---

What it can do

Core

• Chat like a normal AI assistant
• Upload PDFs, text, Word, CSV, JSON, HTML, images, and more
• Ask questions about your own documents
• Keep conversations local and private

Multi-Agent Mode

For harder questions, DeskChat can use multiple internal agents:

1. Answerer — produces a first draft answer
2. Critic — reviews it and points out issues
3. Refiner — improves it into a final answer

This gives you more structured, higher-quality responses when needed.

Document Tools

• Generate academic summaries of uploaded papers
• Extract table-of-contents / section structure
• Extract cited authors and references

Privacy & Control

• Runs entirely on your machine
• No cloud AI APIs
• No telemetry or tracking
• Optional web search using DuckDuckGo
• Optional remote access via ngrok (with basic auth)
• Automatic cleanup of old data

---

Architecture (high level)

You → Streamlit UI → Core Logic → Local Models (Ollama)
                           │
                           ├── Vector Store (Chroma)
                           ├── Document Loaders
                           ├── Multi-Agent Pipeline
                           └── Optional Web Search

Everything stays local unless you explicitly enable web search or remote access.

---

Project structure

deskchat/
│
├── app.py                  # Streamlit entry point
├── config.py               # Configuration & constants
├── requirements.txt
├── README.md
│
├── core/                   # All non-UI logic
│   ├── housekeeping.py     # Cleanup tasks
│   ├── session.py          # Session & chat persistence
│   ├── loaders.py          # File loading logic
│   ├── vectorstore.py      # Indexing and retrieval
│   ├── web.py              # DuckDuckGo search
│   ├── refiners.py         # Summaries, TOC, authors
│   ├── agents.py           # Multi-agent logic
│   ├── generation.py       # Single-agent RAG / chat
│   ├── ngrok.py            # Remote access helpers
│   └── utils.py            # Utility helpers
│
├── ui/                     # Streamlit UI components
│   ├── sidebar.py
│   ├── knowledge.py
│   └── chat.py
│
├── data/                   # Uploaded files
└── storage/
    ├── chats/              # Saved chat histories
    └── indexes/            # Vector indexes

---

Requirements

System

• Python 3.10+
• Ollama installed and running
• 16GB RAM recommended (more for large models)
• GPU recommended but optional

Python dependencies

pip install -r requirements.txt

Ollama models

Pull the models DeskChat uses:

ollama pull qwen3:8b
ollama pull qwen2.5:7b
ollama pull gemma2:9b
ollama pull starling-lm:7b-alpha
ollama pull mxbai-embed-large

---

Running DeskChat

streamlit run app.py

Then open:

http://localhost:8501

---

Optional: Remote access

If you want to access DeskChat from your phone or another device:

ngrok config add-authtoken YOUR_TOKEN

Then enable remote access inside the sidebar UI.

Basic authentication is enabled by default so random people cannot access it.

---

How to use it

Upload documents

• Drag & drop files into the upload area
• Click Build / Rebuild index

Chat modes

• RAG — uses your documents as context
• Plain Chat — normal chatbot, no documents
• Multi-Agent — Answerer + Critic + Refiner pipeline

Document tools

For each uploaded file you can:

• Generate a summary
• Generate a table of contents
• Extract cited authors

---

Data retention

By default:

• Chats and indexes older than 10 days are deleted automatically
• You can change this in config.py

---

Privacy

Data type	Leaves your machine?
Chat messages	No
Documents	No
Embeddings	No
Models	No
Web queries	Only if enabled
Remote access	Only if enabled

---

Disclaimer

• DeskChat does not include content moderation unless you add it.
• Output quality and bias depend on the underlying models.
• Intended for personal use, research, and experimentation.

---

Development

Run with live reload:

streamlit run app.py --server.runOnSave true

License

MIT — do whatever you want with it.