Parrot - Privacy-Aware Agentic RAG for Social Media

What problem is being addressed?
Modern AI assistants are powerful, but struggle to work with users' personal social data in a meaningful and privacy-conscious way. Users generate large amounts of information through posts, comments, direct messages, and online interactions, yet existing assistants either can not access this data effectively or require sending sensitive information to external cloud services. This creates major concerns around privacy, security, and personalization.
PARROT addresses this problem by creating a system that allows users to query and analyze their own social media history using Retrieval-Augmented Generation (RAG) while maintaining strong privacy protections. The system focuses on helping users retrieve insights from conversations, interactions, and posts without exposing raw personal information unnecessarily.
What is the project idea and how will it work?
PARROT is an Agentic RAG system designed for social media analytics and personal knowledge retrieval. The system ingests user activities such as posts, feed interactions, comments, and direct messages, then processes and stores them in a graph-based retrieval system using Neo4j and vector embeddings.
The project consists of several major components:
- Data Ingestion Pipeline:
User social data is collected and normalized into a consistent structure. Posts, comments, and direct messages are embedded using the 'intfloat/e5-large-v2' embedding model for semantic search. - Graph + Vector Database:
The system stores relationships between users, posts, comments, and conversations inside Neo4j. Vector indexes enable semantic similarity search while graph traversal supports analytics and relationship-based queries. - Agentic Retrieval Pipeline:
Queries pass through multiple AI agents including a Query Analyzer, Router, Retrieval Nodes, and a Synthesis Node. The router dynamically selects between simple retrieval, advanced retrieval, or analytics-based graph traversal depending on query complexity. - Analytics Features:
Users can ask:- "Who do I message the most?"
- "What topics do I usually discuss?"
- "What are the main themes in my posts?"
- Privacy Protection:
PARROT supports both local and remote LLM pipelines. When using remote models, sensitive information is anonymized before being sent externally, helping preserve user privacy.
The final result is a personalized AI assistant capable of understanding and analyzing a user's social interactions while combining graph databases, vector search, LLMs, and privacy-aware processing into one unified system.