Aida Learning Companion
Agentic AI Learning Companion

1. Design philosophy

The name of Aida (Learning Companion) comes from the acronyms of the four stages of the AIDA learning framework, and also stands for Agentic Intelligent Development Assistant. Unlike traditional question-and-answer AI assistants, Aida is a learning companion with autonomous action capability (Agentic) and can:

• Active perception: analyse the Student's cognitive level in real time (Bloom's Taxonomy) to determine the learning stage
• Autonomous tool invocation: automatically search Course materials and query Learning history according to the conversational context
• Guide rather than tell: Use Socratic Dialogue, never giving the answer directly, but helping Students discover it for themselves through questions
• Cross-session accumulation: build a personal learner model that becomes more attuned to Students over time

2. AIDA learning framework

The AIDA framework divides learning history into four stages, with each stage corresponding to different cognitive activities and guiding strategies:

Stage	Cognitive Activity	Aida's guiding behaviours	Typical student language
Ask Question	Ask a question, express confusion, clarify needs	Help students turn vague confusion into specific questions	'I don't understand...'; 'Why...?'; 'What does this mean?'
Imagine Conceive	Formulate hypotheses, make associations, compare possibilities	Encourage divergent thinking and provide multi-perspective clues	"I think it might be..." "Could it be..."
Design Plan	Organise ideas, list steps, build a framework	Review the completeness of the plan and ask about omitted aspects	"I plan to... first, then..." "The steps are..."
Act Practice	Execute the plan, produce outcomes, reflect on the results	Provides specific feedback to guide reflection and improvement	"I did... and the result was..." "Is that right?"

Automatic phase detection

After each response, Aida analyses the most recent conversation content and uses a lightweight LLM (GPT-5-mini) to determine the student's current AIDA stage. The detection result must reach a confidence level of at least 60% before updating, to avoid frequent fluctuations.

3. Agentic architecture

Aida uses OpenAI function calling to implement agentic behaviour; all tool definitions are centralised in chat_tools.py(Shared the same tool registry with ClassroomGPT and UeduGPTs). When AI determines that additional information is needed, it will autonomously decide which tools to invoke. Currently there are 45 個工具，分為 9 群；下表於每次載入時直接讀取註冊表產生，與程式永遠一致。

課程情境（9 items, requires classroom_id）

讀取當前課程／頻道的資料；AIDA 自由探索未綁課程時回 {"error": "未指定課程"} 屬預期行為

Tools	Purpose	Parameter
get_course_overview	查詢當前課程的總覽資訊，包括：課程基本資料（課名、學期、系所、授課教師、課號、修課期間）、課綱（教學目標與教學內容）、教材列表、即將到期的測驗、問卷、AI 互動作業，以及蘇格拉底對話與模擬辯論主題。	include_materials、include_deadlines、days_ahead
get_bloom_trend	分析學生在當前課程中的 Bloom's Taxonomy 認知層次趨勢。	time_range
get_forum_stats	查詢學生在當前課程討論區的發文統計，包括發文數、回覆數、獲得的按讚數、計分成績等。	N/A
get_worksheets	查詢當前課程的學習單列表，包括學習單標題、截止時間、學生的作答狀態與成績。	N/A
get_class_recordings	查詢當前課程的課堂影音紀錄（螢幕錄製與音訊錄製），列出已公開的影片與錄音。	N/A
get_podcasts	查詢當前課程的 Podcast 節目列表，包括標題、時長、是否有個人化版本。	N/A
get_peer_comparison	匿名比較學生在當前課程中的表現與全班同學的相對位置。	N/A
get_health_summary	查詢學生的健康與生理數據摘要，包括心率、壓力指數、血氧、步數等。	days
search_channel_materials	搜尋本頻道（教師教室或學生 mygpts）已上傳的課程教材內容（PDF / DOCX / PPTX / TXT），也涵蓋 Uedu K12 認證教材匯入此頻道的內容。	query、top_k

學生通用（5 items）

任何場景可用的計算、繪圖、生圖、課綱查詢與技能載入

Tools	Purpose	Parameter
execute_code	前端瀏覽器透過 Pyodide（WebAssembly）執行 Python 程式碼的程式教學與科學計算沙箱，回傳 stdout / stderr，並自動截獲 matplotlib figure 顯示為圖檔。	code、language
calculate_and_plot	單一數學表達式的求值、求解、化簡、微分、積分（sympy 後端）。	expression、operation、x_range
generate_image	MUST CALL when student explicitly asks to draw/generate any image.	prompt、orientation
get_hs_curriculum	查詢台灣 108 課綱（十二年國教）國中或高中的科目清單與學習內容。	school_level、subject、grade、area
load_skill	載入一個已安裝技能（Skill）的完整操作指引。	skill_name

平台知識（6 items）

回答「Uedu 是什麼／有哪些論文」等問題；description 皆標示「切勿憑記憶回答」以防 hallucination

Tools	Purpose	Parameter
list_uedu_team_papers	查詢 Uedu 研究團隊已發表的學術論文清單（教學實踐研究成果），資料來源為 Uedu 平台官方維護的論文資料庫，內容真實可信。	year、category
list_uedu_team_conferences	查詢 Uedu 學術交流的國際研討會清單。	N/A
list_uedu_universities	查詢 Uedu 平台目前收錄的台灣大學清單。	region
get_uedu_platform_intro	查詢 Uedu（優學院）平台本身的介紹資訊，包括願景、子系統、Educational Omics 框架、核心子系統等。	N/A
list_science_experiments	查詢 Uedu Science 科學實驗室的互動模擬實驗清單（物理／化學／生物／地球科學，全部可在瀏覽器直接操作：調參數、看即時圖表、記錄匯出數據，實驗頁並有 AI 實驗助教陪同）。	subject、keyword
list_emi_teaching_methods	查詢 Uedu「實證教學法文章系列」——每篇對應一份近年同儕審查研究與量化效果量（effect size），並附可在 Uedu 平台實作與驗證成效的具體建議。	keyword、group

知識專區讀本（4 items）

跨域通識知識專區的文章、術語表與資料庫

Tools	Purpose	Parameter
search_knowledge_zone	搜尋 Uedu 跨域通識的知識專區文章。	zone、query、category
get_knowledge_article	依文章 slug 取得某知識專區單篇文章的完整內容（回答後需要更詳細內文時用）。	zone、slug
lookup_knowledge_term	查詢某知識專區的術語詞彙表（中英對照 + 白話定義 + 延伸文章）。	zone、query
query_knowledge_database	查詢知識專區的資料庫。	zone、kind、query、filter

環境資料（3 items）

中央氣象署與環境部公開資料（校園／縣市／全台排行）

Tools	Purpose	Parameter
get_school_environment	查詢學校層級的環境資料：氣溫（含日均/最低/最高）、相對濕度、雨量、風速、UV、日照時數、PM2.5、PM10、O3、NO2、SO2、CO、AQI、CH4/NMHC/THC 碳氫類。	period、school_query
get_county_environment	查詢台灣縣市層級的環境資料：氣溫、相對濕度、雨量、風速、PM2.5、PM10、O3、NO2、SO2、AQI。	county、period、start_date、end_date
get_taiwan_environment_ranking	全台縣市排行——依指定指標（AQI、PM2.5、溫度、雨量等）+ 時間範圍，回傳排序後的縣市清單。	metric、period、order、top_k

財經（6 items）

Uedu 優財經的台股／美股資料（公開資料，非投資建議）

Tools	Purpose	Parameter
search_stocks	搜尋台股／美股／指數，依關鍵字（symbol、中文名、英文名）回傳最相符的標的清單。	query、market、limit
get_stock_overview	查詢單一個股／指數的歷史走勢與最新報價。	market、symbol、lookback_days
get_top_movers_stocks	取台股／美股當日漲跌幅排行（不含指數）。	market、direction、limit
get_foreign_institutional_ranking	取台股三大法人買賣超排行。	type、direction、limit
get_market_indices_snapshot	取台股加權指數、美股三大指數（道瓊、那斯達克、S&P 500）、費城半導體指數（^SOX）的最新報價與漲跌幅。	N/A
get_tw_sector_performance	取台股各產業當日平均漲跌幅、家數、上漲/下跌家數、總成交額。	min_companies_per_sector

指導老師媒合（7 items）

銜接 advisor_match；Uedu 認證老師與 OpenAlex 外部論文嚴格區隔

Tools	Purpose	Parameter
list_advisor_research_topics	列出 Uedu 平台上所有 opt-in 教師的熱門研究主題（依教師人數由高到低排序）。	N/A
find_advisors_by_topic	依單一研究主題關鍵字，找出所有提到該主題的 opt-in 教師（不限制 5 位）。	topic、limit
search_advisor_papers	跨所有 opt-in 教師的論文做關鍵字搜尋（標題與摘要）。	query、limit
search_external_papers	搜尋 Uedu 之外的公開學術論文（OpenAlex 資料庫），預設限定亞洲五國的機構作者：台灣、新加坡、日本、韓國、中國（注意：台灣與中國為兩個獨立國家）。	query、countries、limit
list_research_clusters	列出 Uedu 平台依語意自動分群的研究領域地圖。	N/A
get_cluster_advisors	取得某個語意研究群集底下的所有老師清單。	cluster_id
recommend_potential_advisors	當學生表達想找專題指導老師、研究指導教授、論文指導老師、想加入某個研究領域的實驗室時使用。	interests、university、limit

學習目標（3 items）

跨對話追蹤的學習目標（建立需學生確認）

Tools	Purpose	Parameter
propose_learning_goal	為學生建立一份「學習目標提案」，含分步驟計畫（例：12 週備考計畫）。	title、description、target_date、plan_steps
get_my_goals	查詢學生的學習目標與進度（含待確認提案與進行中目標的計畫步驟勾稽狀態）。	N/A
update_goal_progress	學生回報某個進行中目標的進展（如「大綱寫完了」「第三章讀完了」）時呼叫：勾稽已完成的計畫步驟、記錄進度摘要；學生明確說整個目標完成時可標記完成。	goal_id、progress_note、completed_step_indexes、mark_completed

對話歷史（2 items）

搜尋學生本人跨課程的歷史對話

Tools	Purpose	Parameter
search_my_conversations	搜尋學生過去的對話記錄。	query、course_name、time_range
list_my_courses	列出學生已加入的所有課程名稱及其對話數量。	N/A

Tool invocation flow

1. Student message submitted
2. Aida analyses messages and lets the LLM decide autonomously whether to call tools, and which ones to call
3. If tools are needed → execute → feed the results back to the LLM → possibly call again (multi-round tool call) → generate a response
4. If no tool is needed → generate a guided response directly
5. After the response is completed, run Bloom's classification + AIDA stage detection in the background

Strict mode and hallucination safeguards

All tools are configured with "strict": True and list all required fields in full, in line with OpenAI strict function calling requirements. The description of platform knowledge tools is specifically appended with the instruction “Never answer from memory or guesswork”, forcing the LLM to use tools rather than answer directly, and avoiding hallucinated team papers or conference information.

4. Cognitive level tracking

Each Student message is analysed in the background asynchronously by the Bloom's Taxonomy classifier, recording scores for six cognitive levels:

Level	Description	Typical behaviour
Remember	Recall facts and basic concepts	Define, list, identify
Understand	Explain an idea or concept	Classify, describe and explain
Apply	Apply knowledge in new contexts	Execution, implementation, problem-solving
Analyze	Break down information and build connections	Compare, differentiate, organise
Evaluate	Assessment and decision-making	Critique, defend, evaluate
Create	Generate new ideas or works	Design, build, invent

These data are also used by Aida's guidance strategy reference (via the classify_bloom_level tool), and are imported into the Teacher Console for class-level learning analytics.

5. Dialogue facilitation strategies

Aida's core guiding principles:

• Socratic questioning: each response must include at least one guiding question
• Do not give the answer directly: when the Student gets stuck, provide a small clue rather than a full solution
• Keep responses concise: limit to 2–4 paragraphs to avoid information overload
• Respect autonomy: University students are adult learners; communicate in an equal tone

Task planning

Each time a new conversation is opened, Aida automatically generates the following based on the student's historical session summary and current course information:

• Recommended starting stage for AIDA
• This learning objective
• Specific suggested learning steps
• Personalised opening greeting

Session summary

At the end of the conversation, Aida automatically generates a learning summary, including: topics covered, progress through the AIDA stages, the highest Bloom cognitive level reached, and suggested directions to continue exploring next time.

6. How to use

Students can use Aida in two ways, and the Instructor does not need any extra setup:

Method 1: switch in Uedu TA (recommended)

On the existing course discussion page (/mygpts/), the header area has a [TA] [Aida] toggle button:

• TA mode (default): controlled by the instructor System Prompt, directly answers student questions
• Aida mode: after switching, AI behaviour changes to AIDA-framework guidance, Socratic dialogue

After switching, all existing features (image upload, voice input, doodling, file upload) can still be used. Conversation records are stored uniformly in the original Course log, distinguished by the aida_mode flag.

Option 2: Personal Learning Centre

Go to the standalone page via "Aida Learning Companion" in the global sidebar (/aida/):

• Choose a Course or free exploration (not tied to a Course)
• The left sidebar displays the conversation history
• Each conversation includes task planning (learning goals, suggested steps) and an end summary
• Centralised management of cross-course history

Uedu TA vs. Aida Learning Companion

	Uedu TA	Aida Learning Companion
Location	Course AI TA (an extension of the Instructor)	Personal learning companion (support for students)
Behaviour control	Teacher System Prompt	AIDA framework (built into the platform)
Answer style	Answer the question directly	Socratic guidance without giving the answer directly
When to use	Do you have a specific question that needs answering?	Unsure whether you understand it, or want to explore it in depth
Cognitive Tracking	Bloom's analysis (background)	Bloom's analysis + AIDA stage tracking

7. Teacher dashboard

On the left sidebar of the Course Management page, click "Aida Learning Companion" to view usage data:

Usage overview

• Total number of conversations, number of Students used, average number of messages per conversation
• Bloom's cognitive level distribution chart

Individual Student Analysis

• Conversation count, message count, main Bloom level for each Student
• Most recent activity time

Learning alerts

The warning system has two modes, analysed independently:

• TA mode alert: analyse historical conversations of students under Uedu TA mode
• Aida mode warning: analysing student conversations under Aida guided mode

Each alert is labelled with a severity level (high/medium/low) and a source of judgement (AI detection/data analysis), and can be expanded to view the specific grounds and data behind the judgement.

8. Learning alert mechanism

Warnings are triggered in two complementary ways:

AI real-time detection

On Aida's standalone chat page, AI autonomously decides whether an alert needs to be triggered based on the context of the conversation. Trigger scenarios include:

• The student cannot answer the guiding questions over multiple consecutive turns
• Student expresses frustration, helplessness, or an intention to give up
• The Student's question clearly exceeds the scope of the Course and may require Instructor intervention

Data-driven analysis

The system calculates automatically based on objective data, and is suitable for TA mode and Aida mode:

Warning type	Determination conditions	Severity
Bloom cognitive level too low	Recent conversation messages ≥ 5, with more than 80% staying at the remembering or understanding level	Medium–high
Low dialogue participation	At least 3 conversations, and in more than 80% of cases, each time leave after sending only 1-2 messages	Low

9. Research applications

Aida's conversation data may be used for educational research (subject to IRB review); potential research directions include:

Educational Omics Integration

Omics dimension	Data provided by Aida
Cognomics (cognitive processes)	AIDA stage trajectory, Bloom's cognitive level temporal changes
Linguomics	Analysis of the linguistic complexity and question patterns in student messages
Sociomics	Human–computer interaction patterns, tool usage frequency, session duration

Researchable question

• Does the conversion model in the four stages of AIDA relate to learning outcomes?
• Differences in the effects of Socratic guidance across Students at different Bloom cognitive levels
• Accuracy of learning alerts and the effectiveness of early intervention
• How a cross-course personal learner model evolves over time