Source repository · Star if this helps

Chapter 1 Getting to know FastGPT

Chapter guide: A cross-border e-commerce support team wants a bot that can handle after-sales questions—answers grounded in the policy manual and FAQ, ticket lookup when users report repairs, and delivery through DingTalk and their own platform. Building in-house means stitching together a vector store, an orchestration framework, and channel integration as three separate engineering tracks; FastGPT packs that whole chain into one out-of-the-box platform. This chapter is the cognitive foundation: where FastGPT sits and its capability map; a concept map of apps, knowledge bases, workflows, tools, and skills; choosing among the three deployment forms; a three-way comparison with DIY RAG and ragflow; and finishing with a ten-minute first question on the cloud. When you finish this chapter, you can: state clearly what FastGPT is and is not; pick a form for the "售后客服" ("After-sales Support") case and justify it; and circle a reading route with shared terminology.

flowchart LR A["Needs
After-sales Q&A tied to knowledge · ticket lookup for repairs
Reach users via DingTalk and your own platform"] --> B["Chapter 1 Getting to know FastGPT
Positioning · concepts · form choice · first question"] B --> C["What FastGPT is (1.1)
Knowledge base + workflow
+ app orchestration platform"] B --> D["Core concept map (1.2)
App · knowledge base · workflow
Tools · skills"] B --> E["Three forms (1.3)
Cloud · commercial · community
Sized to scale and governance"] B --> F["Three-way comparison (1.4)
DIY RAG · ragflow
Retrieval depth vs orchestration breadth"] B --> G["Ten minutes to first question on cloud (1.5)
Zero-deploy first app
Ask one question yourself"] C --> H["Payoff
Shared terms · justified choice
Read the book with After-sales Support"] D --> H E --> H F --> H G --> H

Chapter overview figure

When you need it

Three typical situations, and three ways to use this chapter.

The boss asks "should we use FastGPT?" and you need reasons for the meeting—read Sections 1.1 and 1.4. Start with the capability map, then the three-way table against DIY and ragflow; most "yes or no" answers sit in that table.

The project chose FastGPT but you do not know cloud, commercial, or community—go straight to the comparison table and stance in Section 1.3. Feature gaps, pricing shape, and fit for each team type each have a row; you can decide in ten minutes.

You want to click through once before committing to the book—jump to Section 1.5, register for cloud, finish a first question in ten minutes; if it feels right, come back to the concept map in Section 1.2 to plan reading.

This chapter has no prerequisites: Section 1.5 only needs a browser (cloud is ready to use; self-hosted deploy and model channels are in Chapter 2).

1.1 What FastGPT is: knowledge base, workflow, and app orchestration in one platform

The official one-liner (v4.17.3 user guide "Quick look at FastGPT", verified 2026-11-01): FastGPT is an AI Agent application development platform based on large language models (LLM), integrating knowledge-base Q&A, visual workflows, Agent orchestration, tool calling, and skill extension, so developers and business users can quickly build dedicated AI applications.

Split the definition and you get a capability map:

Capability block Official one-liner Plain reading Chapters
Knowledge-base Q&A All-round knowledge base: import varied documents with automatic knowledge structuring; smart Q&A with multi-turn context Documents in, answers with evidence out Chapters 3–5
Visual workflow Drag-and-drop UI to build complex business flows with rich functional nodes Business flows on a canvas Chapter 7
App orchestration Agent orchestration: four forms—conversational Agent, workflow, plugin, Agent V2 Different task complexity on different carriers Chapter 6
Tool calling System tools, custom tools, and MCP tools Apps answer and also "act" Chapter 8
Skill extension Versioned "document + code" capability packs, bound to agents, sandbox execution Turn tuned playbooks into reusable team assets Chapter 8
Publish and integrate Fully aligned with the official OpenAI API; one-click enterprise WeChat, official account, Feishu, DingTalk Ship orchestrated capability outward Chapters 9–10

What matters most for the our running example: intelligent data parsing

The official knowledge-base section on data processing deserves attention: intelligently handle complex PDF structure—keep images, tables, and LaTeX formulas; recognize scans; structure content as clear Markdown; support automatic image captioning and indexing so visual content can be understood and retrieved. That is exactly what the After-sales Policy Manual (after-sales-policy.pdf, 50 pages with tables) needs—whether compensation tables keep row/column structure shows up in Chapter 3 at ingest time; the HTML form of the Shipping Time Notice and the spreadsheet form of the Product FAQ (product-faq.xlsx) also fall under "all kinds of documents."

Strength in the chain, not a single point

Look at each piece alone—vector retrieval, workflow engine, channel access—and there are standalone open-source options. FastGPT's strength is not any single technology; it is packing "build knowledge base, orchestrate flow, publish app" into one product: developers and business users collaborate on one UI, config page equals debug page—the hardest gap on the DIY path (Section 1.4). The official three advantages point the same way: simple and flexible like building blocks (drag to implement complex flows), smarter data (import, preprocess, match knowledge, full Q&A automation), open and easy to integrate (standard API without changing source; supports ChatGPT, Claude, DeepSeek, ERNIE Bot, and other mainstream models).

1.2 Core concept map: app, knowledge base, workflow, tool, skill

Five concepts recur in the next twelve chapters. This table gives each a one-line definition and the object in the our running example—skim now, expand in later chapters:

Concept English One-line definition Object in the our running example Chapters
App App User-facing dialog and flow carrier; four forms: simple configuration, workflow, plugin, Agent V2 "售后客服" ("After-sales Support") app: simple configuration in Chapter 6, workflow as main line in Chapter 7 Chapters 6, 7
Knowledge base Knowledge Base RAG carrier; hierarchy knowledge base → collection → data (Chunk) after-sales-kb: three after-sales document types Chapters 3–5
Workflow Workflow Canvas node orchestration; node is the smallest execution unit Classify question → knowledge search → ticket lookup → reply Chapter 7
Tool Tool External capability the model can call: system, custom, MCP HTTP tool node to middle-platform ticket API GET /api/after-sale/tickets?phone={phone} Chapters 7, 8
Skill Skill Versioned "document + code" pack, bound to agent, sandbox execution; init, develop, integrate, publish Turn after-sales support playbooks into a reusable team capability pack Chapter 8

How concepts connect, in one figure:

flowchart LR APP["App
Simple config · workflow
Plugin · Agent V2"] -->|"Retrieval augmentation"| KB["Knowledge base
Collection → data"] APP -->|"Complex business flow"| WF["Workflow
Canvas orchestration
Node as smallest unit"] WF -->|"Execute actions"| TOOL["Tool
System · custom · MCP"] APP -.->|"Bind integration"| SKILL["Skill
Init · develop
Integrate · publish"]

How core concepts reference each other: apps consume knowledge bases and workflows; workflows drive tools; apps can bind skills

Peripheral concepts quick reference

Beyond the five main concepts, later chapters introduce more "sub-concepts." You do not need them now, but knowing where they belong keeps you oriented:

Peripheral concept Belongs to One-line note Chapters
Collection Knowledge base Organizational layer under a knowledge base; usually one file or one source Chapter 3
Data (Chunk) Knowledge base Smallest knowledge unit; direct object of retrieval and citation Chapters 3, 4
Node Workflow Smallest execution unit on the canvas; edges set order Chapter 7
MCP tool Tool External tool set via Model Context Protocol (MCP) Chapter 8
Channel Publish How apps reach end users: share link, DingTalk, Feishu, WeCom, official account Chapter 9
OpenAPI Publish Platform APIs: apps, chat, knowledge bases consumed via API Chapter 10
SignOz / call logs Operations Observability entry in production Chapter 11

Four app forms: from dialog to autonomous planning

The official quick start shows four cases of increasing capability—enterprise email assistant (conversational Agent), civil code Q&A (knowledge base + conversational Agent), content review and rewrite (workflow), smart data analysis (Agent V2)—plus a selection recap table:

App type (quick start) Fit Core capability
Conversational Agent Light Q&A, copy generation, standardized output Prompt, model config, tool calling
Knowledge base + conversational Agent Q&A on materials, policies, statutes, product manuals File import, knowledge retrieval, cite source
Workflow Fixed steps, branches, review flows, automation Node orchestration, deciders, human confirm
Agent V2 Data analysis, complex tasks, multi-step reasoning, dynamic planning Autonomous planning, tools, VM execution

(Transcribed from v4.17.3 quick start "Four types selection recap", verified 2026-11-01.)

Four selection sentences from the same source: light dialog or standardized copy → conversational Agent first; answers must be grounded in existing materials → knowledge base + conversational Agent; fixed flow with branches, human confirm, or automation → workflow; open task, non-fixed steps, autonomous analysis or code execution → Agent V2. The official text also stresses capabilities can be combined—a support assistant answers product questions from the knowledge base, then queries orders via tools. The our running example "售后客服" ("After-sales Support") is the full combo: knowledge base + workflow + ticket tool, landed in Chapter 7.

Two extra judgments reused in Chapters 6 and 7: a task fits workflow when it has "stable steps, clear rules, repeatable execution" (quick start wording); Agent V2 code runs in a VM, isolated from the local environment.

1.3 Three forms compared and how to choose: cloud, commercial, community

Same FastGPT, three ways to consume it. Quick view:

Form Positioning (official) Who it fits How to start
Cloud Official hosted; register and use Quick try and light use; small teams that do not want ops Register mainland fastgpt.cn or international fastgpt.io (accounts not shared)
Commercial Enhanced build on community; exclusive features Teams offering commercial service; enterprise governance Community image + commercial image (License boot); Sealos fully managed or own servers
Community Free; all core features (Agent build, workflow, knowledge base) Individual developers or small delivery teams Self-deploy (Chapter 2); follow open-source license

Feature gaps: what community lacks

Official feature-diff table between community and paid forms; rows relevant to later chapters (transcribed commercial docs, verified 2026-11-01):

Capability Community Commercial / cloud
Workflow orchestration, share link and API, knowledge base, multi-model config, model log dashboard ✅ ✅
App publish security config, third-party publish (Feishu, official account) ❌ ✅
Run log dashboard, app evaluation, workflow/Agent/Skill assisted generation ❌ ✅
Web site sync, third-party KB scheduled sync, file-level KB permission, image KB, index enhancement ❌ ✅
Team space and permissions, multi-tenant and billing, SSO login, admin console ❌ ✅

Cloud vs commercial differ on some enterprise details: custom copyright and SSO are "in design"; admin console "not required"—full matrix in official commercial docs. This table previews form prerequisites: app evaluation (Chapter 13) and web sync, file-level permissions (Chapter 5) sit on paid forms; community readers should check this table before those sections.

Pricing and delivery shape

Pricing transcribed (official commercial docs, verified 2026-11-01, re-check before publish): Sealos fully managed from ¥10,000/month (8C32G, 3-month minimum) or ¥120,000/year; multi-node from ¥22,000/month (32C128G); own-server deploy within 14 days, price on inquiry; professional services (custom dev, maintenance, migration, third-party integration) ¥2,000–3,000/person/day. Official deploy hints: fully managed for teams short on engineers who only care about business; own-server for full control on your hardware; single-node for internal SMB use; HA for external online service (visual monitoring, replicas, load balance, DB auto backup).

Worth knowing for secondary development: you may modify some community code, not commercial images; after customization, version upgrades need your own merge (official FAQ).

Three common misreadings

Form choice for After-sales Support and one open item

Our recommendation: the our running example "售后客服" ("After-sales Support") starts on community—knowledge base, workflow, share link, and API are in community core; Chapter 2 deploy and Chapter 3 build are not blocked by form. Two exceptions upfront: publish to official account/Feishu channels, app evaluation and run log dashboard need commercial or cloud; DingTalk is not listed with "third-party publish (Feishu, official account)" in the official diff table—community availability to be confirmed hands-on in Chapter 9 (on this chapter's review list).

In business terms: personal learning and small internal teams start on community (deploy in Chapter 2); for multi-tenant external service, SSO/team governance/evaluation dashboard, evaluate commercial and cloud; to try only, register cloud and follow Section 1.5 for ten-minute first question.

1.4 Three-way comparison: DIY RAG, ragflow, and FastGPT

After-sales support needs are three beats: documents to knowledge, knowledge to process, process to service. The three paths differ in where they invest:

Dimension DIY (vector store + orchestration + glue) ragflow FastGPT
Main focus Full-chain DIY, every link replaceable Document parsing and retrieval engineering: deep understanding, hybrid retrieval, chunk governance App orchestration out of the box: workflow, four app forms, channels and API
Starting point Code: vector DB and retrieval scripts Dataset: parsing and chunk tuning App: config pages and canvas
Business collaboration All via engineering backlog Richer ops UI, canvas orchestration Developers and business on one platform; config equals debug
External delivery Custom API and channels API and MCP Server Share link, IM channels, OpenAPI, publish as MCP Server
Best when Deep custom retrieval chain, existing platform, strong compliance self-host Main pain is "complex documents → reliable answers" Main pain is "wire knowledge, flow, channels into a shippable app fast"

In business terms: if retrieval and parsing quality is the main pain, choose ragflow; if app orchestration and delivery speed is the main pain, choose FastGPT; if retrieval must be deeply custom or infra fully self-held, go DIY. They are not exclusive—same after-sales domain can split knowledge base and app layer; Chapter 13 of this book and the RAGFlow book share the methods of chunk governance and quality inspection.

When FastGPT is faster

Our recommendation—any one of three reasons warrants serious consideration: business and engineering build apps on one platform (support ops can change prompt and opening line without a dev ticket); flows must hit business systems (after-sales support checks tickets before reply—HTTP tool to middle-platform ticket API); multi-channel delivery (login-free share for the team, DingTalk for users, OpenAPI for the platform, Chapter 9 MCP Server publish). Conversely, if the pain is only retrieval quality, see (→ ragflow book); if every link must be fully self-held, DIY—engineering detail (vector DB choice, retrieval assembly, Agent consumption) is outside this book (→ agentscope-java book).

Same domain, different names: a terminology crosswalk

FastGPT and ragflow are both knowledge-base platforms; same domain, different names—do not trip when cross-reading (compiled for this book; terms follow each official docs):

Concept domain FastGPT ragflow
Knowledge carrier Knowledge base Dataset
Knowledge unit Data (Chunk) Chunk
Orchestration carrier App / workflow Chat assistant / canvas Agent / ingestion pipeline
External exposure OpenAPI / MCP Server API / MCP Server

One chain, three divisions

User asks "How is damage in transit compensated?"
  → Classify as after-sales inquiry
  → Retrieve after-sales-kb (policy manual + product FAQ)
  → HTTP tool queries ticket status for this user
  → AI dialog summarizes policy and ticket status
  → Publish as login-free share link and DingTalk channel

In FastGPT this is "one knowledge base + one workflow app + two channels"; DIY strings vector DB, orchestration framework, API service, channel SDK; ragflow does retrieval well, but channels and business-flow orchestration are not its home court. Selection sentence: see which segment of the chain holds your main pain.

1.5 Quick start: ten minutes to first question on the cloud

Prerequisite: a browser and email—cloud includes model config, no own model channel (self-hosted channels in Chapter 2). All in the browser:

flowchart LR A["Register cloud
fastgpt.cn or fastgpt.io
Accounts not shared"] --> B["New conversational Agent
Workbench → New
Config page two columns"] B --> C["Pick model · write Prompt
Role + output format + rules
Opening line guides first ask"] C --> D["Debug preview first ask
Test three input types: in-corpus
out-of-corpus · vague"] D --> E{"Answer from materials?"} E -->|"Yes"| F["Create KB and link
Upload → parse → ready
Full build in Chapter 3"] E -->|"Not yet"| G["Publish for trial
Share link in Chapter 9"]

Cloud ten-minute main line: register to first question; optional branch link knowledge base

Six actions from register to first question

  1. Register: open fastgpt.cn (mainland) or fastgpt.io (international). Accounts are not shared (cloud docs)—team must register same region and edition;
  2. New app: workbench "New", choose conversational Agent. Name for testing is fine—the formal "售后客服" ("After-sales Support") app is Chapter 6; do not occupy the name here;
  3. Know the config page: left AI config, right debug preview—change and test; this rhythm is basic FastGPT app development;
  4. Pick model: official cases note Chinese expression and stable format adherence; low-risk tasks like email need not use the strongest model (quick start case 1);
  5. Write Prompt: organize as "role + output format + rules"—role fixes identity, format constrains structure, rules bound risk; this triple repeats in all book prompts;
  6. Opening line and first ask: opening tells users what the app does; put [问题内容] in the opening so users can one-click ask (official tip). Then ask the first question in debug preview.

Do not validate first ask only by "does it sound right"—official knowledge-base Q&A advice applies to no-KB play too: test three input types—in-corpus should answer with evidence, out-of-corpus should say cannot confirm, vague should ask for facts. An app that only agrees is not worth shipping.

After first ask: link a knowledge base

For grounded answers, follow quick start case 2: left sidebar knowledge base → new → text dataset → upload local file → parse settings (chunk storage; chunk condition controls tokens per chunk—defaults for quick test; index enhancement for plain text check first two) → data preview check chunk quality → status "ready" → back to app config link KB. After link, the chain becomes user question → knowledge retrieval → model summary—Chapter 1 touches the RAG door here.

Test file suggestion: one or two pages from the After-sales Policy Manual—feel parsing and citation first. Preview chunk quality for three points—abnormal paragraph breaks, title and body in same semantic span, tables and clauses still readable (quick start). Formal after-sales-kb and three document update rhythms are Chapter 3; parse and retrieval tuning Chapter 4.

Effect difference between the two configs is the key split between official case 1 and case 2 (compiled for this book):

Dimension No knowledge base (case 1) Linked knowledge base (case 2)
Answer basis Model memory Knowledge retrieval
Strong at Generation, formatted output Policy, manual, statute Q&A
Key config Prompt triple + tools Parse params + citation and disclaimer
Main risk Private fact hallucination Chunk quality drives evidence quality

Limits and common mistakes

Chapter summary

  1. FastGPT is an LLM-based AI Agent app development platform: knowledge Q&A, visual workflow, Agent orchestration, tools, skills in one product; strength is the full chain in one product, not a single tech point.
  2. Five core concepts (app, knowledge base, workflow, tool, skill) and four app forms (conversational Agent → knowledge base + conversational Agent → workflow → Agent V2); choose by task complexity; capabilities combine.
  3. Three forms: cloud register-and-use; commercial = community image + commercial image, enterprise extras; community free with all core features, self-deploy. Thread-through case starts community; DingTalk community availability pending Chapter 9 hands-on check.
  4. Three-way comparison: retrieval/parsing main pain → ragflow; app orchestration and delivery speed → FastGPT; deep custom or strong compliance self-host → DIY—judgment is which chain segment holds the main pain.
  5. Next: Section 1.5 cloud first question for feel; self-hosted → Chapter 2 community deploy and model channels; Chapter 3 formal after-sales-kb.

Acceptance checklist

You should be able to:

Review list (before publish)

Further reading