Managed Agent
Knowledgator Managed Agent is an autonomous platform agent tailored for dataset engineering, automated data auditing, 5D-quality evaluation, remediation, and fine-tuning preparation.
While the Knowledgator MCP Server provides a stateless JSON-RPC gateway to inference and platform APIs, the Managed Agent executes complex multi-step data pipelines in an isolated execution environment equipped with specialized domain skills.
MCP Gateway vs. Managed Agent
┌────────────────────────────────────────────────────────┐
│ Knowledgator Managed Agent │
│ │
│ CLI Skills: │
│ • dataset-view — Inspect samples with zero noise │
│ • dataset-validate — Strict JSON schema verification │
│ • dataset-probe — 5D Metrics (Balance, Quality...) │
│ • nl-grep — Semantic audit via zero-shot │
│ • dataset-plan/patch — Capped Hybrid quota & patching │
│ • dataset-sync — Stream datasets via MCP Gateway │
└───────────────────────────┬────────────────────────────┘
│ JSON-RPC (Streamable HTTP / SSE)
▼
┌────────────────────────────────────────────────────────┐
│ Knowledgator MCP Gateway (knowledgator-mcp) │
│ • comprehend_inference • extract_inference │
│ • instruct_inference • open_model_inference │
│ • download_dataset • upload_dataset │
└────────────────────────────────────────────────────────┘
🛠️ Key Capabilities & Skills
- Zero Dirty Context: Operates directly on dataset streams on disk (
10k+samples) without loading thousands of raw samples into the LLM context window. - 5D Dataset Diagnostics: Computes mathematical balance (Shannon Evenness), semantic label quality, vocabulary diversity, syntactic complexity, and structural consistency.
- Capped Hybrid Rebalancing: Intelligently mitigates class skew using floor and ceiling quotas without extreme oversampling or data loss.
- Deterministic Patching: Generates reproducible audit manifests and executes offline file transformations without model hallucinations.
📚 Guides in this Section
- Dataset Remediation — Audit, Cleaning & Repair — Detailed walkthrough of 5D diagnostic metrics, GLiClass semantic auditing, Capped Hybrid planning, and the 7-step remediation lifecycle.