DeepSearch is a knowledge-augmented, high-performance deep retrieval and research engine. It uses structured knowledge and LLMs with tools to deliver enterprise Agentic AI search and research. This document walks through using the DeepResearch agent for report-oriented research and the DeepSearch agent for complex question answering.

1. Preparation

Note: The product has separate front-end and back-end projects. Before you start, ensure both are deployed and running. If not, follow the DeepSearch Full Edition installation guide.

1. Get an LLM API key

DeepSearch needs an LLM. Purchase or host a model from a provider. The steps below use Huawei Cloud as an example.

• Open the ModelArts deployment console (online inference).

• Pick a model and click Enable service.

  

• After enabling, open Invocation guide for model details.

  

• Open OpenAI-compatible API and note the API endpoint and model parameter.

• Under API key management, create a key per the console.

Details: Huawei Cloud ModelArts MaaS user guide

---

2. Get a web search / augmentation API key

DeepResearch uses a web augmentation service (e.g. Tavily) to fetch online information.

Tavily: official documentation

---

3. Model configuration

In Model management, click Add model. Fill in Model name, Model ID, API key, Base URL, and Description.

• Model name: Display name; you choose.

• Model ID: Provider’s model id (matches Huawei model parameter).

• API key: From the provider (Huawei API Key).

• Base URL: Provider API base URL (Huawei API address).

• Description: Optional notes.

  

Test the model:

• In model management, click Test on a model.

• Pick a sample prompt.

• Click start; Test succeeded means configuration is OK.

  

If testing fails, double-check endpoint, model id, and key.

---

2. Using the DeepResearch Agent

1. Prerequisites

Open Task space to reach the agent chat UI.

Before using DeepResearch, complete the following.

---

Select the DeepResearch agent

In the input box type @DeepResearch to pick the agent.

The first time (or if setup is incomplete), a configuration dialog appears—use Agent configuration below, then Save configuration.

After saving, the DeepResearch agent is active.

---

Agent configuration

Click Configure next to the DeepResearch agent.

Three tabs group the settings.

Click Save configuration when done.

General

• Human-in-the-loop: When on, the agent may ask for confirmation at key steps.
• Enable provenance: When on, results show sources.
• Number of sections: Max report sections (1–15); higher values mean more detailed planning.

Search

• Search mode

  • Web augmentation: Use a web augmentation engine.
  • Local search (not available yet): Knowledge base only.
  • Hybrid (not available yet): Local + web.

• Search sources

  • Web augmentation engine: Configure engines (iFlytek, Petal AI, Tavily, Google, custom).
  • Local knowledge base (not available yet): Select one or more KBs.

Templates (optional)

To control report structure, use templates.

• Available templates: Pick from the list.
• Import template: Upload new template from a sample.
  • Extract: Auto-extract structure from Markdown, HTML, Word, or PDF.
    
  • Use as template: Use a Markdown file directly as the template.
    

Templates are optional. Uploaded templates or sample reports must be no larger than 50 MB per file. Very large or structurally complex documents may fail to parse, so splitting the file and retrying is recommended.

Model assignment

Assign models per research stage for quality vs cost.

Basic

• General model (required): Default for all stages when advanced slots are empty. Prefer a strong general model.

Advanced (optional)

Per-stage overrides; unset stages use the general model.

• Outline generation: Prefer a strong reasoning model.
• Information gathering: Lightweight model is often enough.
• Report writing: Prefer models strong at code/math for accuracy.

Advanced is optional. All stages need function calling support. For configurable runtime function-call tools (HTTP APIs), see the Developer Guide and API Reference sections on api_tools_config.

---

Choose the chat model

Use the model picker in the top-right of the chat.

You must select a model. Context under 128K may limit long research reports.

---

2. Conversation flow

2.1 Ask a question

Notes

• Type in area 1, send with 2.
• Two modes: HITL (human-in-the-loop) and non-HITL.

2.1.1 HITL mode

Two user turns:

1. You ask; DeepResearch may ask follow-ups.
   • Yellow indicator (1) = waiting for your reply.
   • Type feedback in 2 and send.
     
     
2. After your reply, the run continues.
   • Green (3) = feedback received.
   • 4 shows your message; 5 shows report generation starting.
     
     

2.1.2 Non-HITL mode

One turn: you ask → DeepResearch generates the report directly.

Report generation behaves the same afterward; below uses non-HITL for the UI tour.

---

2.2 Message panel

2.2.1 Regions

1. 1: Your message.
2. 2: Assistant message.
3. 3: Step status: hollow gray = not started; spinner = running; green check = done; red = failed; yellow = stopped manually.
4. 4: Current section title.
5. 5: Section duration d*h*m*s (updates while running, fixed when done).
6. 6: Expand/collapse body.
7. 7: Sub-sections.
8. 8: Expand/collapse sub-sections.

---

2.2.2 Sub-section progress

---

2.2.3 External citations

Citations from web/local search.

Collapsed vs expanded via the toggle.

---

2.2.4 Sub-report cards

Generating: gray, shows time, not clickable.
Done: blue, shows time, click to open the side panel.

---

2.2.5 Final report

Click the final report to open details on the right.

---

2.3 Conversation history

A conversation is a thread with context. In All conversations you can:

1. Collapse/expand the list.
2. Save a conversation.
3. Start a new one.
4. Open a past thread.
5. Delete a thread.

---

3. Final report

DeepSearch merges retrieved content into a report. A final report card appears in the UI.

---

3.1 Open / close full report

Click the card to toggle the detailed view.

Scroll to read.

---

3.2 Citation details

Click numbered citations [1] for snippets, source, confidence, etc. Use web to open the source page.

---

3.3 Download

Use Download to save locally.

Formats: Markdown, HTML, DOCX.

That completes the Deep Research walkthrough. For deeper topics, see the Developer Guide and hands-on tutorials.

---

3. Using the DeepSearch Agent

Deep Search mode is optimized for complex search tasks. To use the Deep Search mode in openJiuwen, follow these steps:

1. Select Deep Search mode

In Task space, select Deep Search mode.

Deep Search Configuration (steps 2 to 4)

2. General settings

In General tab, configure search process settings such as question routing.

Question routing helps save LLM tokens by routing simple questions to a lightweight ReAct agent, and triggering Deep Search only for complex questions. Other options in this section can be used to control search scope and process behavior.

3. Search settings

In Search tab, select knowledge sources for the search process. Supported sources include local knowledge bases and online search via Serper and Jina. Enter API keys in the corresponding configuration areas.

4. Model settings

In Model tab, configure planning and search models. The planning model is used once at the beginning for question parsing, so a strong general-purpose model is recommended.

5. Enter a Deep Search question

Enter your Deep Search question in Task space and send it.

6. Search process starts

Deep Search begins the retrieval process and shows question-parsing outputs, including unknown entities to search and related statistics.

You can click DeepSearch Explorer for a more detailed view of the full Deep Search process.

7. Get the answer

After the search process completes, the system displays the final answer.