Features

Your infrastructure. Your data.

AI Curator is built around a fundamental principle: your data belongs to you. Run it on your Mac. Deploy it to Docker. Ship it to your cloud. It's always your infrastructure — ElGap never hosts it, never sees your data, never phones home.

Security: own it

AI Curator is open source, provided as-is, in early beta. You can read every line of code. You can audit it yourself. You deploy it, you secure it, you own the infrastructure. We provide the software — you provide the operational security.

Universal Data Capture

Three capture modes cover every workflow: live streaming via HTTP API, file import (JSON, JSONL, CSV), and manual entry in the Web UI. If it can send a POST request, it can feed AI Curator. The same captured data can go to fine-tuning or RAG — you decide at export time.

 curl -X POST http://localhost:3333/api/capture \ -H "Content-Type: application/json" \ -d '{ "source": "slack", "records": [{ "instruction": "How do I reset my password?", "output": "Go to Settings → Account → Reset Password...", "category": "support", "qualityRating": 4 }] }'

No sample ships without your approval

Every sample goes through your review — whether it's training data or knowledge base content. Rate quality. Categorize by domain. Reject what doesn't meet your standards. Only approved data makes it to your training set or your RAG index.

Export for fine-tuning or RAG

Seven export formats optimized for different training pipelines and RAG indexing. Same curation, different destinations.

All exports support filtering by status, quality rating, category, and tags. Train/test/validation splits with stratification are built in. JSONL and CSV exports work for both fine-tuning pipelines and RAG indexing pipelines.

REST API for Curation

The Live Capture API streams data in. The REST API manages curation out. Together, they close the loop: review, rate, approve, and reject — programmatically. Connect external tools, automation scripts, or AI agents to reduce human involvement in the curation loop.

 import requests BASE = "http://localhost:3333/api" # Get pending samples with quality rating samples = requests.get(f"{BASE}/datasets/1/samples", params={ "status": "draft", "minQuality": 3 }).json() # Auto-approve quality 4+ with categories for sample in samples: if sample["qualityRating"] >= 4 and sample.get("category"): requests.post(f"{BASE}/samples/{sample['id']}/approve") else: requests.patch(f"{BASE}/samples/{sample['id']}", json={ "status": "in_review", "tags": ["needs-review"] })

CLI & SDK

When you need to move fast and work at scale. Automate imports, exports, and dataset management from the terminal.

 # Bulk import curator import massive-dataset.jsonl --dataset 1 --workers 8 # Filtered export for fine-tuning curator export --dataset 3 --format mlx \ --filter "status=approved AND quality>=4" # Export for RAG indexing curator export --dataset 5 --format jsonl \ --filter "status=approved AND category=docs" --output rag-corpus.jsonl # HuggingFace search & download curator search "python programming" curator download hf:openai/summarize_from_feedback --dataset 3 # Split for training curator export --split "0.8,0.1,0.1" --seed 42 --format jsonl