DeepRefine-Skill

Type /deeprefine in your AI coding assistant after you've built a graphify knowledge base — it patches graphify-out/graph.json from your session's query history to evolve your LLM-Wiki.

Works in Cursor (install the skill once with deeprefine cursor install). The same workflow is available from any terminal via the deeprefine CLI (deeprefine-cli on PyPI).

/deeprefine

Typical flow: graphify . → graphify query "..." → record the question → /deeprefine.

deeprefine history add --query "your question"   # or let the agent do this

That's it. Under graphify-out/.deeprefine/ you get:

graphify-out/
├── graph.json                          updated graph (graphify reads this)
└── .deeprefine/
    ├── history.jsonl                   queries queued for refinement
    ├── refinement_results_*.jsonl      run logs
    └── graph.json.bak                  backup before each refine

Standalone repo. Model code (autorefiner, atlas_rag) lives in a separate DeepRefine checkout.
pip install deeprefine-cli ships the CLI and SKILL.md. deeprefine refine still needs DeepRefine + atlastune + inference (today: local vLLM).

---

Roadmap

Inference today: local vLLM only (see Inference below).

Coming soon: third-party API support (OpenAI-compatible hosted LLM + embeddings) — no local GPU required.

---

Quick start

Step	What
1	Install DeepRefine in atlastune
2	pip install deeprefine-cli
3	deeprefine cursor install in your KB project
4	Start vLLM (embedding + refine)
5	deeprefine history add → deeprefine refine

# 1) DeepRefine (once)
conda activate atlastune
cd /path/to/DeepRefine && pip install -e .

# 2) CLI (once per env)
pip install deeprefine-cli

# 3) Cursor skill (KB project root)
cd /path/to/your-kb-project
deeprefine cursor install

# 4) vLLM (each session, from DeepRefine repo)
bash /path/to/DeepRefine/scripts/vllm_serve/qwen3-0.6b-emb.sh
bash /path/to/DeepRefine/scripts/vllm_serve/qwen3-8b-vllm-reafiner.sh

# 5) Refine
deeprefine history add --query "your question"
deeprefine refine

---

Pipeline

Stage	Tool	Input	Output
Build	graphify	Project files	graphify-out/graph.json
Query	graphify	Questions	graphify query "..."
Refine	deeprefine	Graph + query history	Updated graph.json, logs

  project files
        │
        ▼ graphify
   graph.json ◄────────────────────────┐
        │                              │
        ▼ graphify query               │ deeprefine refine
   (Q&A session)                      │
        │                              │
        ▼ deeprefine history add       │
   history.jsonl ──────────────────────┘
        │
        ▼ graphify query (verify)

DeepRefine does not build the graph; it patches graph.json so later graphify query retrieves better.

---

Repository layout

DeepRefine-Skill/              ← this repo (PyPI: deeprefine-cli)
├── deeprefine_skill/          ← package (SKILL.md bundled)
└── scripts/deeprefine.py

DeepRefine/                    ← separate clone
├── autorefiner/
├── AutoSchemaKG/
└── scripts/vllm_serve/

your-kb-project/
└── graphify-out/
    ├── graph.json
    └── .deeprefine/           ← history, logs, FAISS cache

Recommended sibling layout (auto-detects ../DeepRefine when DEEPREFINE_REPO is unset):

www/code/
├── DeepRefine/
└── DeepRefine-Skill/

---

Installation

1. DeepRefine (atlastune)

See DeepRefine — Environment.

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

2. CLI

Method	Command
PyPI (recommended)	pip install deeprefine-cli
Source	pip install -e /path/to/DeepRefine-Skill

deeprefine --help    # verify

3. DeepRefine path (optional)

Only if DeepRefine is not ../DeepRefine and not found by walking up from cwd:

export DEEPREFINE_REPO=/path/to/DeepRefine

4. Inference (vLLM)

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. Cursor skill

Run at KB project root (folder with or that will have graphify-out/):

Command	Scope
deeprefine cursor install	.cursor/skills/ (this project)
deeprefine cursor install --user	~/.cursor/skills/ (all projects)
deeprefine install	alias for cursor install
deeprefine cursor uninstall	remove skill

---

Workflow with graphify

One-time

pip install graphifyy deeprefine-cli    # deeprefine-cli in atlastune

cd /path/to/your-kb-project
graphify cursor install
deeprefine cursor install

Each session (KB project root)

#	Action
1	graphify . or /graphify . → graphify-out/graph.json
2	graphify query "..."
3	deeprefine history add --query "..."
4	deeprefine refine or /deeprefine
5	(optional) graphify query "..." to verify

---

Commands

All commands below run from KB project root.

Command	Description
deeprefine history add --query "..."	Record a query after graph Q&A
deeprefine history list	List all history entries
deeprefine history list --pending	List unrefined queries only
deeprefine refine	Refine all pending queries
deeprefine refine --query "..."	Refine one query (also recorded)
deeprefine refine --rebuild-index	Rebuild FAISS before refine
deeprefine index --rebuild	Rebuild FAISS cache only
deeprefine cursor install | uninstall	Manage Cursor skill

Artifacts (graphify-out/.deeprefine/)

File	Purpose
history.jsonl	Query history
refinement_results_*.jsonl	Refinement logs
graph.json.bak	Backup before refine
reafiner.pkl	FAISS index cache

Cursor agent instructions: SKILL.md → installed as .cursor/skills/deeprefine/SKILL.md.

---

Where to run what

What	Where
pip install deeprefine-cli	Anywhere (atlastune for refine)
pip install -e .../DeepRefine	DeepRefine repo
graphify / deeprefine cursor install	KB project root
deeprefine refine / history	KB project root
vLLM serve scripts	DeepRefine repo

---

License

MIT — see LICENSE.