buddy_report.md

Buddy - Indic Language AI Assistant

1. AI Assistant Overview

Assistant Name: Buddy
Purpose & Target Audience:
Buddy is a friendly, multilingual student assistant designed to help Indian students in Telugu, Hindi, and English. Its primary aim is to make access to academic help, doubt solving, and general knowledge easier for Indic language speakers.

Key Features:

• Multilingual support: Telugu, Hindi, and English
• Simple, polite responses tailored for students
• Runs efficiently in a local/low-resource setup
• Feedback capture for continuous improvement
• Downloadable chat history
• Support for factual Q&A and general conversation

---

2. System Prompt Design and Justification

Chosen Open-Source LLM & Environment

LLM: ai4bharat/IndicBART
Deployment/Interaction Environment:
We use a local Python + Streamlit setup. The transformers library loads the model for real-time inference. This ensures the app is free to run without API call costs, and it is truly open-source in both model and hosting.

Full System Prompt

Prompt Justification & Impact

• Persona: The assistant is named “Buddy” to create a student-friendly, non-intimidating experience.
• Tone: Polite and simple answers ensure accessibility for younger students and non-native English speakers.
• Language Handling: The prompt ensures support for three major Indian languages. IndicBART handles translation and generation reasonably well.
• Constraints: No long or verbose responses. Keeps clarity as priority.

Iteration:

We modified decoding logic:

• Use beam search for short queries to reduce randomness.
• Allow sampling for longer or open-ended queries.

This change significantly reduced hallucinations like repetitive or nonsensical answers to factual questions.

---

3. User Reviews and Feedback Analysis

Feedback Collection Plan:
We collected user feedback through a feedback form inside the app (feedback.csv).

User ID	Date	Language	Purpose	Rating	Comment
student_01	2025-08-02	Hindi	GK question	4	“Good answer, slightly slow”
student_02	2025-08-02	Telugu	Basic doubt	5	“Understood everything!”

Summary:

• Accuracy: 80% of queries were correctly answered
• Clarity: Very high for simple questions
• Tone: Polite and respectful
• Issues Found: Struggled with basic math and string repetition in factual queries like “What is 2 + 3?”

Key Improvements Identified:

1. Improve math/factual grounding
2. Add fallback messages for unknown queries
3. Add a clearer language switch indicator

---

4. Future Roadmap

Short-Term Goals (1 Week)

• Add math parsing for small arithmetic
• Improve clarity on language selection
• Polish the UI further with icons and speech-to-text

Mid-Term Goals (2–4 Weeks)

• Add new Indic languages (e.g., Bengali, Marathi)
• Add RAG capability using faiss or custom datasets
• Allow voice input using whisper or VOSK

Long-Term Vision (1+ Month)

• Make Buddy usable by rural students via mobile
• Build an offline APK using Streamlit’s mobile compatibility or Toga/Beeware
• Create learning pathways in Indian languages using Buddy

---

5. Plan to Increase User Adoption

Acquisition:

• Share public repo on GitHub and Hugging Face Spaces
• Promote in regional student forums and Telegram groups
• Present to local colleges as free learning help

Promotion:

• Tag as “Indic Learning Assistant” on GitHub
• Write blogs/tutorials on building open-source chatbots

Feedback Loops:

• In-built feedback with CSV logging
• Future integration with Google Sheets for faster aggregation

Community Growth:

• Accept community language patches via GitHub
• Encourage user testing in different dialects
• Collaborate with NGOs supporting education in regional languages

---

License

This project is released under the MIT License, encouraging reuse, improvement, and educational access.

---

Deployment

To run locally:

git clone https://github.com/<your-username>/buddy-assistant.git
cd buddy-assistant
pip install -r requirements.txt
streamlit run app.py