Documentation / v1 preview

Build a private research loop that keeps its context.

ProjectNew is a local-first workspace for planning research, searching references, running experiments, and turning results into papers without sending your working data to the cloud.

Start setup Explore ProjectNew

Before you begin

Requirements

Python 3.11 or newer

R Engine for R-based workflows

Minimum Hardware (CPU only): 16GB RAM

Minimum Hardware (Small GPU): 8GB RAM + 4GB VRAM

Storage for model files and project data

First launch

Installation

Install ProjectNew, then create a predictable model directory so the application can locate reasoning and embedding models.

C:\model\
├── llm\
│   └── Granite-4.1-3b\
│       └── granite-4.1-3b-Q4_K_M.gguf
└── embedding\
    └── granite-embedding-97M-multilingual-r2\
        └── granite-embedding-97M-multilingual-r2-Q8_0.gguf

1Install ProjectNew and launch the application.
2Place a GGUF reasoning model inside the llm directory.
3Add an embedding model when you need references and retrieval.
4Open Model Configurator and select both files by their full paths.

Get started

From installation to your first local research session.

Prepare the environment

Install ProjectNew

Python 3.11+
R Engine required for R workflows
Minimum 16GB RAM (CPU only) or 8GB RAM + 4GB VRAM (Small GPU)
Storage for model files and project data

Install ProjectNew and launch the application. ProjectNew runs models locally. Minimum requirements: 16GB RAM for CPU-only systems, or 8GB RAM + 4GB VRAM for systems with a small dedicated GPU.

Create the local workspace

Set up your model directory

Create the root model directory
Create a dedicated folder for LLMs
Create a dedicated folder for embeddings

C:\model\
├── llm\
└── embedding\

This structure keeps local inference files and embedding assets organized from the first setup step.

Start with a Model

Download a local LLM

Example: IBM Granite 4.1 3B
File: granite-4.1-3b-Q4_K_M.gguf
Use Q4_K_M as a practical default for a first local setup

C:\model\llm\Granite-4.1-3b\
└── granite-4.1-3b-Q4_K_M.gguf

A practical default quantization for getting started. Users can choose another quantization according to available RAM/VRAM.

Optional for RAG workflows

Add an embedding model

LLM → reasoning, chat, generation
Embedding → references, retrieval, and project knowledge
Example: granite-embedding-97M-multilingual-r2-Q8_0.gguf
Required for advanced retrieval workflows, optional for a basic first launch

C:\model\
├── llm\
│   └── Granite-4.1-3b\
│       └── granite-4.1-3b-Q4_K_M.gguf\
└── embedding\
    └── granite-embedding-97M-multilingual-r2\
        └── granite-embedding-97M-multilingual-r2-Q8_0.gguf

For example, a practical embedding file is granite-embedding-97M-multilingual-r2-Q8_0.gguf from the mykor Granite embedding GGUF repository.

Home → New Project

Create your first project

Project name
Project location
Project description
Project structure / template

This should be the first guided project setup flow for new users rather than a hidden menu or a manual-only explanation.

Research structure

Create your first phase and iteration

Phase = a larger research stage
Iteration = a focused experiment or work cycle
Start with Phase 1 → Iteration 1 for the first project

Project\
└── Phase 1\
    └── Iteration 1

For the first session, keep it simple: create Phase 1 and Iteration 1, then expand later as the project grows.

Open Model Configurator

Load Granite

Context Size
GPU Layers
Temperature
Top P
System Prompt
Embedding configuration for references and retrieval

Model Configurator
LLM Model: C:\model\llm\Granite-4.1-3b\granite-4.1-3b-Q4_K_M.gguf
Embedding Model: C:\model\embedding\granite-embedding-97M-multilingual-r2\granite-embedding-97M-multilingual-r2-Q8_0.gguf

Once the model is loaded, the interface should clearly show a ready state so users know the setup is complete.

Open the first workspace

Start working

Chat Dashboard for project-wide context using Granite as the local LLM
Iteration Dashboard for experiment tasks
References for knowledge and RAG inputs using the Granite embedding model
Paper Editor for LaTeX writing
Diary for research notes and decisions

At this point the user has crossed from installation into a working research loop with both the reasoning model and the embedding model available.

This is the golden path: install → model → project → phase/iteration → ready.

Core workflow

Workspace flow

Create a project

Choose a name, location, description, and starting template.

Add a phase

Use a phase for a larger research stage or a meaningful body of work.

Start an iteration

Turn a question or experiment into a focused, repeatable work cycle.

Connect your knowledge

Add references, notes, code, and papers so the project retains its context.

Local inference

Model configuration

LLM model

Handles reasoning, chat, generation, and tool-based work. Granite, Gemma, and Qwen GGUF families are supported.

Embedding model

Creates searchable representations for references, project knowledge, and retrieval-augmented workflows.

Start with a moderate context size and Q4_K_M quantization. Increase context or GPU layers only after confirming the model runs reliably on your hardware.

Common fixes

Troubleshooting

The model does not load

Confirm the file is a GGUF model and that the configured path points to the file itself.

Responses are slow

Reduce context size or GPU layers, then verify that the model fits comfortably in available memory.

Retrieval is empty

Check that an embedding model is configured and that references have finished indexing.