ConvoCareAI is a multi-agent customer support chatbot designed to reduce response times and improve service quality by leveraging advanced AI technologies including Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), and multi-agent workflows. This system automates and enhances customer interactions in the telecom domain, providing accurate and context-aware assistance on queries such as FAQs, SIM activation, troubleshooting, recharge plans, telecom policies, helpline numbers, and Airtel store locations.
- Multi-Agent Workflow: Orchestrated using LangChain and LangGraph to efficiently route and manage diverse customer queries.
- Retrieval-Augmented Generation (RAG): Enables contextually relevant and up-to-date responses by integrating external data sources.
- Real-Time Data Integration: Uses APIs like Tavily and Serper for fetching live policy updates and store locator information.
- Embedding Storage & Search: Utilizes Qdrant vector database for efficient embedding storage and fast similarity searches.
- Natural Language Processing: Employs SpaCy and FastText for preprocessing, language detection, and text understanding.
- File-based Query Handling: Integrates PyTesseract OCR to extract information from images/documents.
- User-Friendly Interface: Built using Streamlit to provide an interactive chatbot experience.
- Hybrid Work Model: Adapted to a mix of remote and on-site collaboration due to office constraints.
- Programming Language: Python
- AI & NLP: Large Language Models (LLMs), LangChain, LangGraph, HuggingFace Embeddings (all-MiniLM-L6-v2), SpaCy, FastText, PyTesseract
- Databases & APIs: Qdrant Vector Database, Tavily API, Serper API
- Framework: Streamlit for UI
- Geospatial Data: Geopy
- Input Query Processing: Language detection and preprocessing using FastText and SpaCy.
- Multi-Agent Orchestration: Different AI agents handle specific query types using LangChain with LangGraph.
- RAG Component: Queries first trigger retrieval of relevant documents/data from Airtel datasets (CSV, PDF, TXT) and live APIs.
- Response Generation: LLMs generate natural language responses based on retrieved context.
- Output: Interactive and dynamic replies served to users through the Streamlit interface.
- Developed expertise in multi-agent AI system design and deployment.
- Successfully integrated large-scale telecom datasets with real-time APIs to provide comprehensive support.
- Tackled challenges in query routing, language identification, and accurate information retrieval.
- Balanced simultaneous tasks of data preprocessing, agent development, and UI deployment within a fixed timeline.
- Collaborated effectively across data science, customer experience, and IT teams in a hybrid work environment.
The project was supported by knowledge gained in the following courses:
- Artificial Intelligence (CSE3705)
- Generative AI and LLMs (CSE3720)
- Generative AI Agents – Task Automation with LLM Reasoning (CSE3024)
- Expand the chatbot to support additional languages and regional dialects.
- Integrate voice-based queries and multimodal inputs.
- Enhance agents with adaptive learning from user feedback.
- Deploy in a cloud environment for scalable access.