Deep Research
multi-step research cycle

1. Design philosophy

Deep Research is a research mode on the Uedu platform for learning questions that need 'cross-referencing multiple data sources and multi-step reasoning'. When a student asks a question involving course material integration, cross-disciplinary comparison, literature search, course planning, or other issues that are hard to answer in one sentence, AI automatically enters a ReAct (Reasoning + Acting) cycle, first planning, then calling tools step by step to gather evidence, continuously reflecting on whether the evidence is sufficient, and finally integrating it into a complete answer with citations.

This design addresses two limitations of ordinary LLM conversations:

• A single Q&A round is hard to span multiple data sources: when a student asks “What is the teaching sequence for comparing list / tuple / dict / set?”, it is necessary to refer simultaneously to the Course materials, teaching papers and semester progress—one LLM call is not likely to integrate across data sources autonomously
• Lack of verifiable citations: model-generated answers directly are prone to hallucination, and Students cannot check the source

Deep Research is off by default. Only after the teacher enables it via the switch in section 6 of Channel Settings, 'Deep Research', will the student chat box show a teal magnifying-glass button.

2. ReAct loop

Each research task follows the Plan → (Action → Observation → Reflect)^* → Synthesize cycle until the LLM determines that sufficient data has been collected, or a system limit is reached.

Subproblem decomposition (Planner)

Planner splits the original question into 2–6 sub-questions that can be researched independently, and suggests possible tools for each sub-question. The complexity judgement affects the number of subsequent execution steps and the overall token budget:

Real-time SSE streaming

All stages are pushed to the front end via Server-Sent Events. Event types include connected, task_created, status_change, plan_complete, action_thinking, action_dispatch, action_complete, reflection_done, limit_hit, cancelled, completed, failed and the heartbeat sent every 15 seconds. The front end can display steps, tokens and tool-call status in real time.

3. Applicability judgement (Auto-judge)

Once Deep Research is activated, it consumes 30-60 seconds and tens of thousands of tokens, so it is not worth enabling for every question. The system provides two trigger paths:

Forced Trigger

After a Student explicitly clicks the magnifying-glass button and submits a question, the engine is triggered directly with trigger_mode='forced'.

Auto automatic suggestions

When a student enters a question in the general input box, the front end uses a lightweight LLM to determine whether it is worth activating. The auto-judge examines the following features:

The classifier also returns a confidence score (0~1). When the confidence score is above the threshold, the front end shows a card saying "This question is suitable for Deep Research"; the Student can decide whether to upgrade to DR mode or keep the general conversation.

4. Tool orchestration

The Action stage of Deep Research shares the chat_tools.py tool registry used by ClassroomGPT / Aida, but the subset actually called is dynamically adjusted according to the question context. Common tool-orchestration combinations:

Channel knowledge base integration

search_channel_materials This is the tool most often called in course contexts for Deep Research. It uses semantic retrieval (cosine similarity) on the RAG materials for this channel uploaded by the Instructor (PDF / DOCX / PPTX) to identify relevant passages, and returns the file name, page number and original passage for Synthesizer citation. See detailsRAG Methodology。

External paper search

search_external_papers uses the OpenAlex public API to query academic papers from institutions in five countries: Taiwan, Singapore, Japan, South Korea and China. The engine strictly separates the Synthesize stage:

• Uedu certified instructor papers (from search_advisor_papers): can label author and provide an advisor profile link
• External paper (from OpenAlex): marked as “external data”, no “apply” or “contact” entry is provided, and it does not imply the author is willing to accept new students

5. Integration and citation

Synthesiser design

Observations from all Action stages are sent together with the original question to the Synthesizer LLM, which outputs structured JSON (answer + citations). The Synthesizer uses response_format={"type":"json_object"} to enforce JSON output, and instructs the LLM to cite sources in the answer with labels such as [1] and [2].

Parameter tuning

The Synthesiser is the most error-prone stage in the entire process, because it has to integrate a large amount of observations within a fixed token budget. In practice, two pitfalls have been encountered:

• reasoning_effort='medium' will consume most of the max_completion_tokens, leaving too little space for the output and resulting in an empty string
• max_completion_tokens set too low (e.g. 6,000) easily causes truncation part-way through the output

At present, the production environment is configured as reasoning_effort='low' + max_completion_tokens=12000, and finish_reason is checked for length to detect truncation.

6. Five layers of cost and security safeguards

Deep Research is the platform's most expensive single feature. To avoid misuse, abuse or loss of control leading to runaway costs, the engine has built five layers of defence:

Cancellation billing policy

• Student-initiated cancellation: deduct 0.5 time(s) (encourages stopping proactively when the direction is found to be wrong, to avoid waste)
• System forced stop (when any one limit is reached: 30 steps / 200K tokens / 10 minutes): deduct 1.0 time (an answer has already been produced and still counts towards usage)

7. Quotas and billing

Role determination

At startup, the engine determines the user role according to the following priority order and sets today's quota accordingly:

1. Platform admin → unlimited
2. Channel owners (including student-created mygpts) → Unlimited (teacher)
3. Channel teacher table (classroomgpt_teacher) → Unlimited (teacher)
4. TA table (classroomgpt_ta) → 40 times / day (ta)
5. No — 20 times / day (student)

Quota window

Uses a rolling 24-hour window, counted 24 hours forward from the time of first activation. Each successful activation deducts 1.0 use (deductions on cancellation follow the rules).

Use API

• POST /api/deep_research/start — Start research (SSE stream)
• POST /api/deep_research/cancel — Cancel an in-progress research task
• GET /api/deep_research/quota — Check current remaining quota
• GET /api/deep_research/my_tasks — Personal history list
• GET /api/deep_research/task/<uuid> — Single task details

8. Persistence and sharing

Three-tier persistence

The dual-row design of Chat logs serves two purposes: (1) answers appear in the dialogue history and remain visible after refreshing the page; (2) the LLM context of subsequent conversations carries over the research findings, so students can continue to ask in-depth questions based on the results.

Consistency between real-time and refreshed data

When research starts, the engine first writes the request log to classroomgpt_my_log (and triggers Bloom cognitive level analysis), then pushes request_log_id via SSE to the front end, after which the front end polls the Bloom badge for that question bubble. At the end of the research, it writes the response log and replaces the inline progress card in the conversation history in place with the answer bubble. After refreshing the page, it reloads from chat history, and the visual result matches the live rendering.

Share Link

The results page for each research task is at /deep_research/task/<uuid>, and can be shared with peers or instructors. Access is restricted: only the owner, channel owner, Super TA, and platform admin can view it, to prevent sensitive research content (including students' personal questions) from leaking.

Personal history page

/deep_research/my Lists all research records of the user themself (including completed, cancelled, and failed), filterable by status and number of items; each item can be opened to review the full answer.

9. Ethics and permissions

IRB ethics foundation

The student usage data for Deep Research (task, steps, answer) are all covered by Uedu's existing IRB approval (NTU-REC 202507EM058). When teachers enable the DR switch on the channel, the default is to follow that informed-consent framework — using the channel is treated as consent for conversation data to be included in research analysis.

Instructor responsibilities

• Assess whether the channel is suitable before enabling it (semester length, whether research-based learning is required)
• Monitor DR usage within the channel (available from admin / channel dashboard)
• Remind Students: the DR answer is an initial result in which AI integrates multiple data sources; it must not be used directly for assignment submission, and cited content must be checked by the Student

Data boundary

The DR toolkit strictly follows Uedu platform data boundary rules:

• Will not read RAG materials from other channels (search_channel_materials is limited to the current channel)
• Will not read sensitive data such as students’ grades, forum posts or AI conversation histories
• External papers (OpenAlex) are strictly separated from Uedu-certified teachers, so as not to mislead students into "applying" to non opt-in instructors

10. Research applications

Educational Omics Integration

Data generated by Deep Research can span multiple dimensions of the EO framework, and is one of the few features that can provide cognitive, language, and social learning-behaviour signals at the same time:

Researchable question

• The relationship between Deep Research activation rates and learning outcomes across different disciplines / learning stages
• The correlation between the quality of subproblem decomposition (Planner) and the quality of the final answer
• Auto-judge accuracy for applicable judgements (adoption rate vs. Student final decision)
• The timing of Student DR cancellation and its relationship with subsequent learning behaviour
• How the ratio of Uedu teaching materials versus external papers in the citation list changes by year level

Data export

Researchers may follow Uedu's existing data governance process (sign the researcher responsibility form, export the consent form, super TA NDA) to apply to export DR-related data. SeeData export and data governance methodology。