How It Works
How It Works
Section titled “How It Works”Plugy is an AI support platform that processes every customer message through a structured cognitive cycle and continuously improves through self-learning.
The 4-Step Cognitive Cycle
Section titled “The 4-Step Cognitive Cycle”Every incoming message passes through four steps:
| Step | Name | What Happens |
|---|---|---|
| 1 | SCAN | Load conversation history, customer profile, and session context |
| 2 | THINK | Analyze intent, emotion, tone, and urgency using a fast AI model |
| 3 | RETRIEVE | Find relevant answers from your knowledge base using semantic search |
| 4 | RESPOND | Generate a context-aware response using the primary AI model |
After the response is sent, the platform updates customer profiles, computes quality scores, and feeds data into the self-learning system.
Multi-LLM Engine
Section titled “Multi-LLM Engine”Plugy uses multiple AI models for different tasks:
- Fast model (e.g., Gemini Flash) — Used in the THINK step for quick intent analysis
- Primary model (e.g., GPT-4) — Used in the RESPOND step for high-quality answers
- Backup model — Automatic failover if the primary model is unavailable
You can configure which models to use per project in the dashboard.
Knowledge Base (RAG)
Section titled “Knowledge Base (RAG)”Your bot’s knowledge comes from your uploaded documents, FAQs, and rules. Plugy uses semantic search to find the most relevant knowledge chunks for each customer question.
- Upload PDFs, text files, or enter FAQs directly
- Automatic embedding and indexing
- Real-time retrieval during conversations
B-Score: Quality Metric
Section titled “B-Score: Quality Metric”Every bot response is scored on four dimensions:
| Component | Measures |
|---|---|
| Focus | Is the answer relevant to the customer’s question? |
| Empathy | Does the tone match the situation? |
| Consistency | Does it align with previous answers and knowledge base? |
| Experience | Does it use conversation history effectively? |
The B-score is a weighted combination of these four components. If any component is zero, the entire score collapses — ensuring no dimension is neglected.
Self-Learning (ONDA)
Section titled “Self-Learning (ONDA)”The ONDA self-learning system continuously improves your bot by:
- Analyzing conversation quality metrics
- Proposing improvements to the bot’s persona and knowledge base
- Testing changes against a benchmark dialog corpus
- Applying only improvements that pass safety checks
Learn more in the Self-Learning section.
Channels
Section titled “Channels”Plugy connects to multiple messaging platforms. Customers can reach your bot through Telegram, JIVO, and more — all managed from one dashboard with shared knowledge and consistent responses.
See Channels for the full list.