{Jig}

Structured output pairings for local + multimodal LLMs (LM Studio · Ollama)

🚀 Quick Start 🧭 Workflow 🖼️ Example 🤝 Contribute

Jig turns natural-language briefs into JSON Schema + system prompt pairings, then runs deterministic inference against your local models. It highlights the multimodal zai-org/GLM-4.6 checkpoint while remaining compatible with every model you can host in LM Studio or Ollama (Llama, DeepSeek, Mistral, etc.).

Jig exists for repeatable LLM inference: it first co-designs the schema and system prompt you need, then guarantees that every future run follows that contract—perfect for production ETL, reporting, or any workflow where ad-hoc prompts are too brittle.

---

📑 Table of Contents

• 📢 News
• ✨ Highlights
• 🖼️ Example: Incident/Urgency Triage in Jig
• 🚀 Quick Start
• 🧭 Workflow in Detail
• 🧠 Local Model Matrix
• 📁 Pairings Layout
• 🛠️ CLI Reference
• 🤝 Contributing
• 🔐 Security
• 📄 License

📢 News

• 2026-02-05 – Added new CLI/Docs, inline pairing previews, and the manual editor tab.
• 2026-02-04 – Inference tab now previews the selected pairing (schema + prompt).

---

✨ Highlights

• Pairing lifecycle – Author schemas with the Creator, preview/test them in Inference, then fine-tune every field inside the Editor.
• Recommended model – Pull zai-org/GLM-4.6 once in Ollama/LM Studio for a single text+vision checkpoint; swap to Llama 3, DeepSeek-VL, Phi, etc. whenever you like.
• Structured enforcement – JSON Schema (Draft 7) is applied via LM Studio's response_format or Ollama's format, so outputs never drift.
• Vision ready – Drag/drop multiple images into the Image tab; Jig handles base64 plumbing for GLM-4.6, Llava, Moondream, DeepSeek-VL, and friends.
• Local-first – No cloud calls. Everything happens against your LM Studio or Ollama runtime.

Without Jig	With Jig
Prompt drift across runs	Schema-locked outputs
Hand-written parsing logic	Auto validation + JSON Schema
No visibility into changes	Versioned pairings on disk

---

🖼️ Example: Incident/Urgency Triage in Jig

This walkthrough shows how the three Gradio tabs work together to build and run a real pairing.

Purpose

“Tag incidents with urgency (critical/high/medium/low) and impacted system so on-call engineers can instantly see what needs attention and where.”

Step 1 — Creator Tab

Describe the task in plain English and Jig generates the schema + system prompt.

jig create "Tag incidents with urgency (critical/high/medium/low) and list the impacted system" -n auto_triage

Generated schema (auto_triage/schema.json):

{
  "type": "object",
  "description": "The root object representing an incident tag.",
  "properties": {
    "urgency": {
      "type": "string",
      "description": "The urgency level of the incident. Must be one of 'critical', 'high', 'medium', or 'low'.",
      "enum": ["critical", "high", "medium", "low"]
    },
    "impacted_system": {
      "type": "string",
      "description": "The system that is impacted by this incident."
    }
  },
  "required": ["urgency", "impacted_system"],
  "additionalProperties": false
}

Generated system prompt (auto_triage/prompt.txt):

You are an expert in creating precise, well-structured JSON Schemas. Your task is to generate a JSON Schema that defines the structure for tagging incidents with two key pieces of information: urgency level (critical/high/medium/low) and impacted system. The schema must adhere strictly to the specified rules: use strict typing with additionalProperties set to false, ensure all properties are required, include clear descriptions for each field, and follow JSON Schema Draft 7.

Step 2 — Inference Tab

Select any pairing, review the schema/prompt preview on the right, and run deterministic inference with text, files, or images.

jig run -s auto_triage -i "Users are reporting that they can’t complete checkout on the web app. Payments fail with a 502 error when submitting the order. The issue seems isolated to the payments service. Mobile app appears unaffected for now. Revenue impact likely if not resolved quickly."

Output

{
  "urgency": "critical",
  "impacted_system": "payments"
}

Step 3 — Editor Tab

Manually tweak schemas or prompts, reformat JSON, and save the pairing—all without leaving the UI.

Bonus — Use the pairing from Python

from jig import SchemaAgent, create_client

client = create_client(backend="lmstudio")  # or backend="ollama"
agent = SchemaAgent(client)

incident = (
    "Users report checkout failures (502 errors) on the web app. "
    "Payments service is impacted; mobile unaffected. Revenue at risk."
)

result = agent.run(incident, "auto_triage")
print(result)

---

🚀 Quick Start

Follow this step-by-step guide to go from idea to deterministic outputs in a few minutes.

Step 1 — Install Jig and extras

# Fastest: install directly from GitHub
pip install git+https://github.com/Leodotpy/jig.git

# Or clone for local development
git clone https://github.com/Leodotpy/jig.git
cd jig
pip install -e .
pip install -e ".[gradio]"  # optional UI extras

Step 2 — Ready your runtime

Runtime	What to do
LM Studio	Load zai-org/GLM-4.6 (or any other model), enable the local server under the Developer tab (default port 1234).
Ollama	ollama serve → ollama pull zai-org/GLM-4.6 (multi-modal). Still works with llama3, deepseek:*, etc. Default port 11434/11435.

Tip

If LM Studio isn't running, --backend auto will transparently fall back to Ollama.

Step 3 — Create a pairing

Use natural language to tell Jig what you want; it will craft the system prompt + JSON Schema pairing for repeatable runs.

jig create "Extract meeting details: attendees, decisions, follow-ups" -n meeting

Step 4 — Run deterministic inference

Feed the model any text (and optional images/files); Jig enforces the schema it generated.

jig run -s meeting -i "Meeting on Feb 3: Alice assigned budget analysis." -o result.json

Step 5 — Launch the Gradio UI (optional)

jig --gradio

• Creator tab – Generate fresh pairings from a plain-English brief.
• Inference tab – Preview schema & prompt before you hit "Run Inference" (text/file/image inputs supported).
• Editor tab – Pick any pairing (or type a new name) and hand-edit schema + prompt with live validation.

---

🧭 Workflow in Detail

1. Creator
   • Describe the task → Jig returns response_schema, system_prompt, and metadata stored under pairings/<name>/.
   • Force overwrite with the checkbox, or let auto backups protect existing work.
2. Inference
   • Select a pairing; the right rail mirrors its JSON Schema & prompt so you can sanity check before execution.
   • Mix modalities: text area, .txt upload, and multi-image gallery all funnel into the same run.
   • Stream results or save directly to disk.
3. Editor
   • Dropdown (with custom values) lets you load any pairing or start blank.
   • Schema formatter button pretty-prints + validates JSON before saving.
   • Saving writes schema.json, prompt.txt, and updates meta.json descriptions instantly.

---

🧠 Local Model Matrix

Capability	Suggested model	Notes
Text + Vision (default)	zai-org/GLM-4.6	Single checkpoint handles both modalities; works in LM Studio and Ollama.
Text-only	llama3, deepseek-r1, phi-4, etc.	Use when you only need deterministic text extraction.
Vision-focused	deepseek-vl, llava, moondream	Great for OCR, UI parsing, etc. Images upload via Gradio Image tab or CLI --image.

Jig targets the OpenAI-compatible (LM Studio) or Ollama-native APIs, so any local model that those runtimes expose is fair game.

---

📁 Pairings Layout

pairings/
├── meeting/
│   ├── schema.json   # JSON Schema enforced at inference time
│   ├── prompt.txt    # System prompt / instructions
│   └── meta.json     # Description, model provenance, timestamps
├── invoice/
│   └── ...

Manual edits? Use the Editor tab or edit these files directly—Jig will pick them up instantly. Overwrites create timestamped backups (*_backup_YYYYMMDD_HHMMSS).

---

🛠️ CLI Reference

Command	Purpose
jig create ...	Turn a natural-language description into a pairing.
jig run ...	Execute inference with text / file / --image inputs.
jig list	Show all pairings with completion status.
jig show <name>	Print schema + prompt contents in the terminal.
jig models [--set <name>]	List models discovered via the current backend or change the active one.
jig --gradio	Launch the Gradio UI (Creator · Inference · Editor).

Common flags: --backend lmstudio|ollama|auto, --model <id>, --temperature, --image path, --output result.json.

---

🤝 Contributing

We welcome improvements in the classic GitHub workflow:

1. Fork & clone this repository.
2. Create a feature branch (git checkout -b feat/my-change).
3. Install dev extras: pip install -e ".[dev]".
4. Format with Black: run black . (VS Code users can enable “Format on Save” with Black).
5. Lint & test: ruff check . and pytest.
6. Open a pull request describing the change and referencing any issues.

Bug reports or feature suggestions via GitHub Issues are equally appreciated.

Roadmap ideas

• Multi-step agent workflows that chain multiple pairings.
• Support for audio-capable LLMs.

---

🔐 Security

See SECURITY.md for coordinated disclosure guidelines.

---

📄 License

MIT – see LICENSE.