Content Summarization Flow

Trigger: Student requests summary via chat interface
Duration: 3–8 seconds
Example: "Summarize topic: Recurrent Neural Networks"

---

Process Overview

The summarization flow uses retrieval-augmented generation (RAG) to create personalized summaries based on the student's course materials.

Student Query → Qdrant Search → Context Building → LLM Synthesis → Response

---

Detailed Steps

1. Query Input

Student submits a natural language request through the chat interface:

• Context provided:
  • course_id - Student's enrolled course
  • student_id - For personalization
  • language - Response language (e.g., pt-PT)

2. Context Building

The system prepares the search:

• Build Qdrant filter: course_id == student's course
• Optional: Narrow by section_id or module_id if specified
• Extract topic keywords from student query

3. Semantic Search

Find relevant course materials:

{
  "collection": "course_materials",
  "query_vector": "<embedded student query>",
  "filter": { "course_id": 123 },
  "top_k": 5,
  "threshold": 0.7
}

4. Result Evaluation

Scenario	Action
k < 3 results	Return: "Insufficient materials; please consult instructor"
k >= 3 results	Proceed to summarization

5. Prompt Construction

The LLM receives a structured prompt:

System Prompt:

You are an educational assistant for Portuguese-speaking university students. 
Summarize the following course materials clearly and concisely.

Structure your response as:
1. **Definição** – What is this concept?
2. **Conceitos principais** – Key points to understand
3. **Exemplos práticos** – Real-world or course examples
4. **Aplicações** – How is this used?

Language: Portuguese (Portugal)
Tone: Educational, accessible
Length: 200–400 words

User Prompt:

Student's request: {query}

Course materials retrieved:
{chunk_1}
{chunk_2}
{chunk_3}
{chunk_4}
{chunk_5}

6. LLM Summarization

Generation parameters:

Parameter	Value	Rationale
Model	GPT-4 or Claude 3	High quality output
Temperature	0.5	Balance consistency and creativity
Max tokens	600	Sufficient for structured response

7. Response Delivery

The response is formatted for readability:

• Format: Markdown with headings and bullet points
• Citation: "Based on materials from: [filename, section]"
• Quality logging: Similarity scores recorded for QA

---

Example Response

Student Query

"Summarize topic: Redes Neurais Recorrentes"

Sample Response:

Definição

Redes Neurais Recorrentes (RNNs) são um tipo de rede neural artificial projetada para processar sequências de dados, mantendo uma "memória" interna de informações anteriores.

Conceitos Principais

• Estados ocultos: Armazenam informação das entradas anteriores
• Backpropagation Through Time (BPTT): Algoritmo de treino específico
• Vanishing gradient: Problema comum em sequências longas

Exemplos Práticos

• Previsão de texto (autocompletar)
• Tradução automática
• Análise de sentimento

Aplicações

RNNs são utilizadas em processamento de linguagem natural, reconhecimento de voz, e análise de séries temporais.

Baseado em materiais de: lecture_05_rnn.pdf, Secção 3

---

Configuration Options

Setting	Default	Description
top_k	5	Number of chunks to retrieve
similarity_threshold	0.7	Minimum relevance score
max_response_length	400 words	Summary length limit
temperature	0.5	LLM creativity level

---

Next Steps

• Doubt Clarification - Question answering flow
• Quiz Generation - Adaptive assessments