deeprefine-cli 0.1.3

CLI and Cursor skill to refine graphify knowledge graphs with DeepRefine

DeepRefine-Skill

Cursor skill and CLI to refine graphify knowledge graphs with DeepRefine (graphify-out/graph.json + session query history).

This repository is standalone. The DeepRefine model code (autorefiner, atlas_rag) lives in a separate DeepRefine checkout.

---

Repository layout

DeepRefine-Skill/          ← this repo (pip install -e .)
├── README.md
├── SKILL.md
├── pyproject.toml
├── deeprefine_skill/
└── scripts/deeprefine.py

DeepRefine/                ← separate clone (training + Reafiner)
├── autorefiner/
├── AutoSchemaKG/
└── ...

your-kb-project/           ← your data (graphify-out/)
└── graphify-out/graph.json

Recommended clone layout:

www/code/
├── DeepRefine/
└── DeepRefine-Skill/      # sibling → auto-detected if DEEPREFINE_REPO unset

---

How graphify and DeepRefine fit together

Stage	Tool	Input	Output
Build	graphify	Project files	graphify-out/graph.json, report, HTML
Query	graphify	Questions	graphify query "..."
Refine	DeepRefine (this repo)	Graph + query history	Updated graph.json, logs

DeepRefine does not build the graph; it edits graph.json incrementally so later graphify query works better.

---

Setup

1. DeepRefine environment (atlastune)

Follow DeepRefine/README Environment to create atlastune and install the main repo:

conda activate atlastune
cd /path/to/DeepRefine
pip install -e .

2. Install this CLI

From PyPI:

conda activate atlastune
pip install deeprefine-cli

From source (development):

pip install -e /path/to/DeepRefine-Skill

Verify:

deeprefine --help

deeprefine refine still requires a local DeepRefine checkout (autorefiner, atlas_rag) and running vLLM services — see below.

3. Point to DeepRefine (if not cloned as sibling)

Only needed when DeepRefine is not at ../DeepRefine relative to this repo and not found by walking up from your cwd:

export DEEPREFINE_REPO=/path/to/DeepRefine

Add to ~/.bashrc if you use a fixed path.

4. Start vLLM (before deeprefine refine)

From the DeepRefine repo:

conda activate atlastune
bash /path/to/DeepRefine/scripts/vllm_serve/qwen3-0.6b-emb.sh
bash /path/to/DeepRefine/scripts/vllm_serve/qwen3-8b-vllm-reafiner.sh

Variable	Default
DEEPREFINE_LLM_URL	http://127.0.0.1:8134/v1
DEEPREFINE_EMBED_URL	http://127.0.0.1:8128/v1
DEEPREFINE_MODEL	HaoyuHuang2/DeepRefine-v1-8B
DEEPREFINE_EMBED_MODEL	Qwen/Qwen3-Embedding-0.6B

5. Install Cursor skill (KB project root)

cd /path/to/your-kb-project    # must contain or will contain graphify-out/
deeprefine cursor install

Same pattern as graphify cursor install. Use --user for all projects.

---

Workflow with graphify

One-time

pip install graphifyy
graphify cursor install          # in KB project
pip install -e /path/to/DeepRefine-Skill   # in atlastune
deeprefine cursor install        # in KB project

Per session (KB project root)

1. /graphify . or graphify . → graphify-out/graph.json
2. graphify query "..." or /graphify query "..."
3. deeprefine history add --query "..."
4. /deeprefine or deeprefine refine
5. Optional: graphify query "..." again to verify

project files ──graphify──► graph.json
                                 │
                       graphify query
                                 │
                 deeprefine history add
                                 │
                 deeprefine refine
                                 │
                       graphify query

---

Command cheat sheet (KB project root)

deeprefine history add --query "..."
deeprefine history list --pending
deeprefine refine
deeprefine refine --query "..."
deeprefine index --rebuild

---

Where to run commands

Command	Directory
pip install -e .../DeepRefine-Skill	Any
pip install -e .../DeepRefine	DeepRefine repo
graphify / deeprefine cursor install	KB project root
deeprefine refine	KB project root
vLLM serve scripts	DeepRefine repo

Agent instructions: SKILL.md.

---

Publish to PyPI (maintainers)

cd /path/to/DeepRefine-Skill
python -m pip install --upgrade build twine
python -m build
twine check dist/*
twine upload dist/*    # needs PyPI token: TWINE_USERNAME=__token__ TWINE_PASSWORD=pypi-...

Test install from TestPyPI first (optional):

twine upload --repository testpypi dist/*
pip install -i https://test.pypi.org/simple/ deeprefine-cli