Core Concepts

FastGPT is built upon a clean domain model. Understanding how applications, knowledge assets, external extensions, and access boundaries interact helps you build resilient, maintainable enterprise AI systems.

System Topology Overview

FastGPT organizes operational entities into configuration, runtime, and persistent storage layers:

Application Entity Hierarchy

In FastGPT, an Application is the primary unit delivering a complete conversational or procedural experience to end users and external systems.

Application Types

• Chat Agent: A conversational application configured around a single chat model, answering queries through system instructions, context history, and attached knowledge bases.
• Chat Agent V2: An advanced agent equipped with dynamic planning and tool invocation capabilities, using a ReAct loop to evaluate goals, select tools, and perform multi-step reasoning.
• Workflow: A deterministic execution pipeline orchestrated with nodes and directed edges on a visual canvas, governing data transformation, conditional branching, and batch execution.

Lifecycle and Snapshots

• Draft Version: An active workspace configuration where builders edit and test changes. The platform supports automatic draft updates and explicit manual checkpoints.
• Published Version: A validated baseline frozen for production use. Publishing synchronizes the configuration to the team and updates the active application version across all publishing channels.
• Publishing Channel: An external ingress point exposing application functionality, including the Login-Free Window, embedded chat widgets, and OpenAPI keys. MCP services are configured separately from the "MCP Services" entry in the Studio sidebar.

Knowledge Asset System

A Knowledge Base is a structured repository of source material prepared through an automated ingestion pipeline, providing verifiable ground truth for Retrieval-Augmented Generation (RAG).

Three-Tier Organization

1. Knowledge Base: The top-level administrative boundary defining global embedding models, text search settings, and collaborator access control.
2. Collection: The organizational unit within a Dataset, displayed in the interface as Files (Collections) and Folders. Files store ingested documents, web pages, or blank collections; Folders organize nested items.
3. Data Chunk: The granular, independently maintainable unit of information produced by segmentation. Each chunk links to one or more searchable indexes.

Dual-Engine Indexing

FastGPT pairs semantic vector retrieval with keyword-based full-text indexing:

• Embedding Index: Encodes semantic features into dense vector representations, excelling at semantic similarity and synonymous expressions.
• Full-Text Index: Employs tokenization and BM25 ranking for exact keyword matches, ensuring high recall for proper nouns, serial numbers, and specialized codes.
• Reranker: Evaluates multi-channel candidates through a unified relevance scoring pass, delivering the most accurate contextual chunks to the model.

Tool and Extension Ecosystem

FastGPT interfaces with external platforms and custom code through modular extension standards:

• Tool: A callable operation with explicit input and output schemas. Related operations are organized under a Tool Set with shared configuration and connection credentials.
• Workflow Tool: A standalone workflow with declared inputs and outputs that can run independently or be mounted onto an Agent as a sub-routine.
• Skill: A versioned bundle containing instructions, templates, and script assets, loaded into execution sandboxes to supply specialized procedural guidance to agents.
• MCP Tool: An external tool exposed to FastGPT through the standardized Model Context Protocol.
• MCP Service: A published FastGPT gateway that presents internal applications and knowledge bases as callable tools to external clients.

Security and Permission Context

The system enforces strict access control across organizations and resources:

• Team Management: Establishes tenant boundaries. Team members are Owner, Administrator, or Common members.
• Resource Collaborator Permissions (ACL): Controls granular access levels (Read permission, Write permission, Administrator) on individual applications and knowledge bases; for applications the read level is shown as "Chat only", and "View chat logs" can be granted separately.
• Platform Administration: Managed by platform root administrators, governing global model catalogs, tenant resource quotas, and system infrastructure.
• Identity Proxy: An authorization mode where an external integration executes requests on behalf of a designated active team member, resolving access rules using that member's credentials.

Next Steps

• Selection guide: Read Choosing Application Types.
• Hands-on tutorial: Start building in Quick Start Guide.
• Knowledge management: Explore Knowledge Base Lifecycle.