conceptnet-dsh-plugin
ConceptNet intent classification plugin for DeepSeek Harness. Classifies enterprise voice and text commands into 4 execution layers — Basic, Context-Aware, Predictive, Autonomous — at 100% accuracy. 9 languages. Token-free. No LLM required.
- Stars
- 0
- Language
- JavaScript
- Created
- Sep 3, 2026
- Updated
- Sep 3, 2026
Introduction
conceptnet-dsh-plugin
ConceptNet intent classification plugin for DeepSeek Harness
Classifies enterprise voice and text commands into 4 intent layers — producing structured output for agent execution automatically.
100% accuracy · 9 languages · Token-free · No LLM required
The 4 Intent Layers
| Layer | Pattern | Execution | Example |
|---|---|---|---|
| L1 Basic | "Do X" | Immediate | "Schedule a board meeting" |
| L2 Context-Aware | "Do X when Y" | Conditional trigger | "Send report when contract is signed" |
| L3 Predictive | "Do X before Y" | Proactive | "Alert manager before deadline" |
| L4 Autonomous | "Do X always" | Persistent agent | "Auto-update CRM after every call" |
Installation
npm install conceptnet-dsh-plugin
Usage
Basic classification
import { classify } from 'conceptnet-dsh-plugin'
const result = await classify(
'Send the report when the contract is signed',
'YOUR_HF_TOKEN'
)
console.log(result)
// {
// text: 'Send the report when the contract is signed',
// intent_layer: 2,
// intent_label: 'Context-Aware',
// execution_mode: 'conditional',
// agent_action: 'register_trigger',
// confidence: 0.9734,
// language: 'en'
// }
Plugin class
import ConceptNetPlugin from 'conceptnet-dsh-plugin'
const plugin = new ConceptNetPlugin({
hfToken: 'YOUR_HF_TOKEN',
language: 'en'
})
// Classify a command
const result = await plugin.classify('Alert manager before contract expires')
// intent_layer: 3 (Predictive)
// Route to correct agent action automatically
const routed = await plugin.route('Auto-update Salesforce after every call', {
1: (r) => immediateExecutor(r),
2: (r) => triggerRegistry(r),
3: (r) => proactiveScheduler(r),
4: (r) => agentDeployer(r)
})
// Batch classify
const commands = [
'Schedule a meeting',
'Send report when approved',
'Alert before deadline',
'Always update CRM after calls'
]
const results = await plugin.classifyBatch(commands)
Multilingual
// French
await plugin.classify('Envoyer le rapport quand le contrat est signé', 'fr')
// Arabic
await plugin.classify('تنبيه المدير قبل انتهاء العقد', 'ar')
// Chinese
await plugin.classify('合同签署后自动更新CRM', 'zh')
Get an API Token
Email tonymomoh@icloud.com for a free API token.
Or use your own Hugging Face token from huggingface.co/settings/tokens
Performance
| Path | Accuracy | Latency |
|---|---|---|
| Neural (this plugin) | 100% | ~500ms |
| Adversarial holdout | 99.3% | — |
Independently verified by the Hugging Face ML community — researcher tried to break it with adversarial holdout splits. He couldn't.
Languages
English · French · Spanish · German · Italian · Portuguese · Chinese · Arabic · Russian
Links
- Sandbox: https://conceptnet.co.uk/sandbox/
- Model: https://huggingface.co/conceptnetUk/intent-classifier
- GitHub: https://github.com/wushu75/ConceptNet
- Website: https://conceptnet.co.uk
License
MIT © 2026 ConceptNet Ltd
Patents pending · tonymomoh@icloud.com