Key Concepts
Understanding these core concepts will help you get the most out of the AI Platform.
Application Types
The platform supports four types of AI applications, each suited to different scenarios:
| Type | Description | Best For |
|---|---|---|
| My Assistant | Conversational Q&A application with multi-turn dialogue | Customer service, knowledge Q&A, business assistants |
| Workflow | Single-run automated task pipeline | Document processing, content generation, data transformation |
| Conversational Workflow | Multi-turn dialogue with visual process orchestration | Complex business processes, guided task execution |
| Translation Assistant | Specialized text translation with glossary management | Document translation, terminology-consistent localization |
Nodes
Nodes are the building blocks of workflows and conversational workflows. Each node performs a specific function, and nodes are connected together on a canvas to form a complete process.
Common node types:
- Start — Entry point; receives user input and injects system variables
- Model — Calls a large language model to process input and generate output
- Knowledge Base — Retrieves relevant text segments from a private knowledge base
- Condition — Routes the flow to different branches based on variable values
- Code Execution — Runs custom Python or NodeJS logic
- HTTP Request — Calls external APIs or services
- End / Direct Reply — Terminates the flow and outputs the final result
Variables
Variables carry data between nodes. There are several variable types:
| Type | Description |
|---|---|
| System Variables | Automatically injected by the platform (e.g., sys.query, sys.user_id) |
| Input Variables | Defined on the Start node; collected from the user before the flow runs |
| Node Output Variables | Produced by each node after execution; referenced by downstream nodes |
| Session Variables | Persist across turns in a conversational workflow (Chatflow only) |
Knowledge Base
The knowledge base is a RAG (Retrieval-Augmented Generation) module. You upload documents, the platform chunks and vectorizes them, and at runtime the system retrieves the most relevant segments to inject into the model's context — enabling accurate, document-grounded answers.
Key settings:
- Embedding Model — Converts text to vectors for semantic search
- Retrieval Mode — Vector search, full-text search, or hybrid
- Top K — Number of segments to return
- **Score Thresholimum similarity score to include a segment
- Rerank Model — Re-ranks retrieved results for higher precision
Prompt
A prompt is the instruction given to a large language model to define its behavior, role, and output format. Prompts can reference upstream node variables using syntax, allowing dynamic, context-aware instructions.
Components
Components are reusable capabilities that can be plugged into applications. They include:
- Built-in components — Platform-provided integrations (e.g., Google Search, Bing)
- Custom components — User-defined HTTP services described via OpenAPI Schema
- Published workflows — A workflow published as a component, callable from other flows
Knowledge Recycling
Knowledge Recycling is a feedback loop that turns high-quality conversation answers into reusable knowledge. Users thumbs-up good answers during conversations; the platform collects them nightly, and administrators review and approve them into a QA knowledge base that improves future responses.
Glossary
The glossary module controls how specific terms are translated in the Translation Assistant:
- Corpus — Maps source terms to fixed transls (e.g., "药理学" → "pharmacology")
- Desensitization Glossary — Replaces sensitive content with placeholders before sending to the model, then restores the original after translation