Quick Start

From registration to your first knowledge base — 30 minutes

What this tutorial will take you through

This tutorial will take you through the basic process of ZGI quickly: prepare materials, create knowledge base, configure large models, orchestrate agents, and publish them into a usable AI application.

Once completed, you will have an agent that answers questions based on your company profile. For example, after uploading contracts, systems, manuals, and FAQs, you can directly ask questions:

• “What is the start and end time of this contract?”
• “What is the reimbursement process in the system?”
• “What are the payment periods and liquidated damages rules for a certain contract?”
• “What contracts are about to expire in the past three months?”

With this tutorial, you will accomplish 4 things:

• Import data and create a knowledge base

• Configuration of available large models

• Build a basic knowledge base question and answer agent

• Publish as an external application, that is, the smart agent landing page

Suggestions

For your first experience, start with a small number of high-quality documents, such as 3 to 5 contracts, policies, product manuals, or FAQs. First run through the basic process, and then gradually increase the amount of data.

Preparation before starting

• A batch of test materials: PDF, Word, Excel, image scans or Markdown documents
• One available source of large models: existing models on the platform, cloud model API keys, or local models in private deployments
• If you use private deployment, you need to prepare the corresponding deployment environment in advance

Build AI application in four steps

Import data and create a knowledge base

Enter the knowledge base module and upload the prepared information. ZGI will automatically complete parsing, slicing, and build a searchable knowledge base.

Capabilities supported by the knowledge base:

• Documents are automatically parsed and sliced
• Vector recall and graph recall
• Rerank
• TopK, confidence threshold, embedding model configuration

It is recommended for the first test: upload a small amount of high-quality documents first → wait for the parsing and slicing to be completed → use the knowledge base test function to verify → check whether the recalled content is accurate → confirm the effect before importing in batches.

Configure available large models

Enter the large model configuration module and select or access the available large models. When trying it for the first time, it is recommended to configure a model that can support basic question and answer to ensure that the knowledge base question and answer process can run smoothly.

In actual use, the large model is mainly responsible for understanding user questions and generating final answers combined with content recalled from the knowledge base.

Orchestrate agent workflow

Enter the workflow canvas and create a basic agent process. The simplest knowledge base question and answer process can be built like this:

Start node → Knowledge retrieval node → LLM node → Direct reply node

• Start Node: Questions that receive user input
• Knowledge retrieval node: Find relevant content from the specified knowledge base
• LLM node: generate answers based on user questions and search results
• Direct reply node: Return the final result to the user

Debug and publish as external application

After the workflow is built, test it in the debugging area first. It is recommended to test at least 3 types of questions:

• Precision query: For example, “What is the expiration time of the contract?”
• Summary type: For example, “Help me summarize the core terms of this contract.”
• Comparative Analysis: For example, “What are the differences in the payment periods of these contracts?”

After confirming that the answer is stable, publish the agent as an external application (agent landing page), and users can access it directly through the landing page.

Developer quick call

In addition to publishing through the interface, developers can also directly call the knowledge base or workflow through the API. A typical request will include the API Key, user question, model configuration and knowledge base identification, and the response content and citation source will be returned.

curl -X POST https://api.zgi.cn/v1/chat \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zgi-default",
    "messages": [
      { "role": "user", "content": "Check which contracts expire in the next three months" }
    ],
    "knowledge_base_id": "kb_xxxxxx"
  }'

Sample response:

{
  "id": "msg_xxxxxx",
  "content": "Based on knowledge base retrieval, the following contracts will expire within three months...",
  "citations": [
    {
      "source": "Procurement_Contract_2025Q4.pdf",
      "page": 3,
      "paragraph": "Article 5 Contract Term..."
    }
  ]
}

Each answer comes with a quote from the original text

The citations field returns specific documents, page numbers and paragraphs, so business teams can verify them directly without guessing the source.

Next step

• Core concepts: Understand the relationship between workspace, agent, knowledge base and model
• Knowledge Base & RAG: Learning slicing, recall, TopK, Rerank and optimization of embedding models
• Agent Orchestration: Build more complex business process applications