CourseTA - AI Teaching Assistant

CourseTA is an Agentic AI-powered teaching assistant that helps educators process educational content, generate questions, create summaries, and build Q&A systems.

Features

• File Upload: Upload PDF documents or audio/video files for automatic text extraction
• Question Generation: Create True/False or Multiple Choice questions from your content
• Content Summarization: Extract main points and generate comprehensive summaries
• Question Answering: Ask questions and get answers specific to your uploaded content

Demo

Requirements

• Python 3.9+
• Dependencies listed in requirements.txt
• FFmpeg (for audio/video processing)
• Ollama (optional, for local LLM support)

Installation

1. Clone this repository:

   1 2	https://github.com/Sh-31/CourseTA.git cd CourseTA

2. Install FFmpeg:

   Linux (Ubuntu/Debian):

   1 2	sudo apt update sudo apt install ffmpeg

3. Install the required Python packages:

   1	pip install -r requirements.txt

4. (Optional) Install Ollama for local LLM support:

   Windows/macOS/Linux:

   • Download and install from https://ollama.ai/
   • Or use the installation script:

   1	curl -fsSL https://ollama.ai/install.sh | sh

   Pull the recommended model:

   1	ollama pull qwen3:4b

5. Set up your environment variables (API keys, etc.) in a .env file.

   Update .env with your credentials:

   1	cp .env.example .env

Usage

Running the Application

1. Start the FastAPI backend:

   1	python main.py

2. In a separate terminal, start the Gradio UI:

   1	python gradio_ui.py

Architecture

CourseTA uses a microservice architecture with agent-based workflows:

• FastAPI backend for API endpoints
• LangChain-based processing pipelines with multi-agent workflows
• LangGraph for LLM orchestration

Agent Graph Architecture

CourseTA implements three main agent graphs, each designed with specific nodes, loops, and reflection mechanisms:

1. Question Generation Graph

The Question Generation agent follows a human-in-the-loop pattern with reflection capabilities:

Nodes:

• QuestionGenerator: Initial question creation from content
• HumanFeedback: Human interaction node with interrupt mechanism
• Router: Decision node that routes based on feedback type
• QuestionRefiner: Automatic refinement using AI feedback
• QuestionRewriter: Manual refinement based on human feedback

Flow:

• Starts with question generation
• Enters human feedback loop with interrupt
• Router decides: save (END), auto (refiner), or feedback (rewriter)
• Both refiner and rewriter loop back to human feedback for continuous improvement

2. Content Summarization Graph

The Summarization agent uses a two-stage approach with iterative refinement:

Nodes:

• SummarizerMainPointNode: Extracts key points and creates table of contents
• SummarizerWriterNode: Generates detailed summary from main points
• UserFeedbackNode: Human review and feedback collection
• SummarizerRewriterNode: Refines summary based on feedback
• Router: Routes to save or continue refinement

Flow:

• Sequential processing: Main Points → Summary Writer → User Feedback
• Feedback loop: Router directs to rewriter or completion
• Rewriter loops back to user feedback for iterative improvement

3. Question Answering Graph

The Q&A agent implements intelligent topic classification and retrieval:

Nodes:

• QuestionClassifier: Analyzes question relevance and retrieves context
• OnTopicRouter: Routes based on question relevance to content
• Retrieve: Fetches relevant document chunks using semantic search
• GenerateAnswer: Creates contextual answers from retrieved content
• OffTopicResponse: Handles questions outside the content scope

Flow:

• Question classification with embedding-based relevance scoring
• Conditional routing: on-topic questions go through retrieval pipeline
• Off-topic questions receive appropriate redirect responses
• No loops - single-pass processing for efficiency

Key Architectural Features

Human-in-the-Loop Design:

• Strategic interrupt points for human feedback
• Continuous refinement loops in generation and summarization
• User control over when to complete or continue refining

Reflection Agent Architecture:

• Feedback incorporation mechanisms
• History tracking for context preservation
• Iterative improvement through dedicated refiner/rewriter nodes

Async API Architecture

CourseTA implements a comprehensive async API architecture that supports both synchronous and streaming responses, providing real-time user experiences and efficient resource utilization.

API Documentation

File Processing APIs

1. Upload File

Upload PDF documents or audio/video files for text extraction and processing.

URL: /upload_file/

Method: POST

Content-Type: multipart/form-data

Request Body:

1

file: Upload file (PDF, audio, or video format)

Response:

1
2
3
4
5
6

{
  "message": "File processed successfully",
  "id": "uuid-string",
  "text_path": "path/to/extracted_text.txt",
  "original_file_path": "path/to/original_file"
}

Supported Formats:

• PDF: .pdf files
• Audio: .mp3, .wav formats
• Video: .mp4, .avi, .mov, .mkv, .flv formats

2. Get Extracted Text

Retrieve the processed text content for a given asset ID.

URL: /get_extracted_text/{asset_id}

Method: GET

Path Parameters:

• asset_id: The unique identifier returned from file upload

Response:

1
2
3
4

{
  "asset_id": "uuid-string",
  "extracted_text": "Full text content..."
}

Question Generation APIs

3. Start Question Generation Session

Generate questions from uploaded content with human-in-the-loop feedback.

URL: /api/v1/graph/qg/start_session

Method: POST

Request Body:

1
2
3
4

{
  "asset_id": "uuid-string",
  "question_type": "T/F"  // or "MCQ"
}

Parameters:

• asset_id: Asset ID from file upload (required)
• question_type: Question type - “T/F” for True/False or “MCQ” for Multiple Choice (required)

Response:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

{
  "thread_id": "uuid-string",
  "status": "interrupted_for_feedback",
  "data_for_feedback": {
    "generated_question": "string",
    "options": ["string"],  // or null
    "answer": "string",
    "explanation": "string",
    "message": "string"
  },
  "current_state": {}
}

4. Provide Question Feedback

Provide feedback to refine generated questions or save the current question.

URL: /api/v1/graph/qg/provide_feedback

Method: POST

Request Body:

1
2
3
4

{
  "thread_id": "uuid-string",
  "feedback": "string"
}

Parameters:

• thread_id: Session ID from start_session (required)
• feedback: Feedback text, “auto” for automatic refinement, or “save” to finish (required)

Response:

1
2
3
4
5

{
  "thread_id": "uuid-string",
  "status": "completed", // or "interrupted_for_feedback"
  "final_state": {}  // or null
}

Content Summarization APIs

5. Start Summarization Session (Streaming)

Generate content summaries with real-time streaming output.

URL: /api/v1/graph/summarizer/start_session_streaming

Method: POST

Content-Type: text/event-stream

Request Body:

1
2
3

{
  "asset_id": "uuid-string"
}

Parameters:

• asset_id: Asset ID from file upload (required)

Streaming Response Events:

1
2
3
4

data: {"thread_id": "uuid", "status": "starting_session"}
data: {"event": "token", "token": "text", "status_update": "main_point_summarizer"}
data: {"event": "token", "token": "text", "status_update": "summarizer_writer"}
data: {"event": "stream_end", "thread_id": "uuid", "status_update": "Stream ended"}

6. Provide Summarization Feedback (Streaming)

Refine summaries based on user feedback with streaming responses.

URL: /api/v1/graph/summarizer/provide_feedback_streaming

Method: POST

Content-Type: text/event-stream

Request Body:

1
2
3
4

{
  "thread_id": "uuid-string",
  "feedback": "string"
}

Parameters:

• thread_id: Session ID from start_session_streaming (required)
• feedback: Feedback text or “save” to finish (required)

Streaming Response Events:

1
2
3

data: {"thread_id": "uuid", "status": "resuming_with_feedback"}
data: {"event": "token", "token": "text", "status_update": "summarizer_rewriter"}
data: {"event": "stream_end", "thread_id": "uuid", "status_update": "Stream ended"}

Question Answering APIs

7. Start Q&A Session (Streaming)

Answer questions based on uploaded content with streaming responses.

URL: /api/v1/graph/qa/start_session_stream

Method: POST

Content-Type: text/event-stream

Request Body:

1
2
3
4

{
  "asset_id": "uuid-string",
  "initial_question": "string"
}

Parameters:

• asset_id: Asset ID from file upload (required)
• initial_question: The first question to ask about the content (required)

Streaming Response Events:

1
2
3

data: {"type": "metadata", "thread_id": "uuid", "asset_id": "uuid"}
data: {"type": "token", "content": "answer text..."}
data: {"type": "complete"}

8. Continue Q&A Conversation (Streaming)

Continue an existing Q&A session with follow-up questions.

URL: /api/v1/graph/qa/continue_conversation_stream

Method: POST

Content-Type: text/event-stream

Request Body:

1
2
3
4

{
  "thread_id": "uuid-string",
  "next_question": "string"
}

Streaming Response Events:

1
2
3

data: {"type": "metadata", "thread_id": "uuid"}
data: {"type": "token", "content": "answer text..."}
data: {"type": "complete"}

Headers for Streaming APIs

Required Headers:

1
2
3

Accept: text/event-stream
Cache-Control: no-cache
Connection: keep-alive