Meshmind

Inspiration Traditional LLMs struggle with large context windows, often hallucinating and losing accuracy when processing multiple documents. We envisioned MeshMind as the "Cursor for data" - an AI-powered system that can intelligently work with any database or document collection, providing accurate answers across multiple files without the limitations of context window constraints.

What it does? MeshMind is an enterprise-grade RAG system that: Processes unlimited files simultaneously (unlike LLMs limited to 3 files) Prevents hallucination through hybrid retrieval with knowledge graphs Works as "AI Drive" - intelligent access to any document database Provides accurate answers with source citations and confidence scores Supports multiple file formats (PDF, TXT, DOCX) with automatic processing Enables conversational AI over your entire document collection

How we built it

Phase 1: Core System FastAPI monolithic architecture with immediate processing PyPDF2 for document extraction and intelligent chunking OpenAI embeddings with hybrid search (vector + keyword) Knowledge graph construction using entity extraction Mock services for rapid development and testing

Phase 2: Enterprise Architecture Microservices separation (ingest/query/worker APIs) MongoDB for document storage, Neo4j for knowledge graphs Asynchronous job processing with Celery Redis caching and session management Production-ready error handling and monitoring

Phase 3: Advanced Integration MCP (Model Context Protocol) implementation External API integration (Notion, Jira) Unified content processing pipeline Enterprise-grade scalability and security

Challenges we ran into Context Window Limitations: Traditional LLMs couldn't handle multiple large documents Hallucination Prevention: Ensuring accurate answers with proper source attribution Scalability: Building a system that works with enterprise-scale document collections Integration Complexity: Connecting multiple databases and external services Performance Optimization: Balancing accuracy with response speed Knowledge Graph Construction: Extracting meaningful entities and relationships from unstructured text

Accomplishments that we're proud of Solved the multi-file problem: Successfully processes unlimited documents simultaneously Eliminated hallucination: Hybrid retrieval ensures accurate, cited answers Built enterprise architecture: Scalable microservices with proper separation of concerns Created knowledge graphs: Intelligent entity extraction and relationship mapping Achieved production readiness: Comprehensive error handling, monitoring, and deployment Delivered MCP integration: Future-proof protocol for external service integration

What we learned Hybrid approaches work best: Combining vector search, keyword matching, and knowledge graphs Architecture matters: Microservices enable better scalability and maintainability Knowledge graphs are powerful: Entity relationships significantly improve retrieval accuracy Mock services accelerate development: Rapid prototyping without external dependencies User experience is crucial: Clear status indicators and error handling improve adoption Documentation is essential: Comprehensive guides enable team collaboration

What's next for MeshMind Real-time collaboration: Multiple users working on the same document collection Advanced analytics: Document insights, usage patterns, and knowledge discovery Mobile applications: iOS and Android apps for document access on-the-go Enterprise integrations: Slack, Microsoft Teams, and other workplace tools AI-powered insights: Automated document summarization and trend analysis Global deployment: Multi-region support with edge computing Advanced security: End-to-end encryption and compliance features API marketplace: Third-party integrations and custom connectors

Vision: MeshMind will become the standard platform for AI-powered document intelligence, making any database or document collection as accessible and intelligent as having a personal AI assistant for your data.

Built With

Submitted to

RowdyHacks XI

Created by

I was responsible for implementing the end-to-end chat session architecture, connecting the frontend, backend, authentication, and database layers into a cohesive workflow. My primary focus was on enabling users to upload files (such as PDFs), start conversations, and have all their messages and documents persist securely across sessions.

On the frontend, I developed an intuitive user interface using React and Material UI (MUI) components. I implemented file selection, progress handling, and API triggers to upload files and send chat messages. Once a file is chosen, the client uploads it to Amazon S3 using a pre-signed URL mechanism, which ensures that large files can be directly uploaded to the cloud without burdening the backend server. After upload, I retrieved the file’s unique metadata (including _id, filename, and S3 URL) and passed it to the chat API for further processing.

For user authentication, I integrated Clerk — enabling secure sign-in and token-based access control. Each API request made from the frontend includes a Bearer token obtained via Clerk, ensuring that only authenticated users can access or modify their chat sessions. This authentication flow ties every message and uploaded file to a specific user identity, providing session-level security and personalization.

On the backend, I designed and implemented the /chat/session endpoint using Express.js, which acts as the intermediary between the frontend and the AI model layer. The Express API handles the creation and updating of chat sessions in MongoDB, storing session-level metadata such as the user ID, message history, timestamps, and the S3 file reference rather than the file itself. I optimized the database logic using $push and $setOnInsert to append new messages instead of overwriting them, preserving the full conversation context. Additionally, I implemented messageCount and updatedAt fields to track session activity and message growth. Once the chat data is validated and stored, Express retrieves the file directly from AWS S3 using its secure URL and then forwards both the file content and conversation context to a FastAPI service, which handles the AI inference tasks such as embedding generation, document retrieval, or LLM-based responses. This architecture creates a smooth and modular bridge between the Node.js backend, AWS S3 for file storage, and the Python-based FastAPI inference layer.

To ensure smooth frontend–backend communication, I implemented robust Axios-based API calls that transmit user messages and file information to the backend, where data is validated, stored, and linked appropriately. This architecture allowed the system to support multi-file chat scenarios, where each user session corresponds to a specific file context.

Overall, my contribution resulted in a fully functional, secure, and scalable chat-with-PDF system that seamlessly integrates file management, authentication, and persistent conversations. The integration of AWS S3, Clerk Auth, MongoDB Atlas, Express API, and MUI frontend demonstrates a complete full-stack implementation, ready for expansion into AI-powered chat workflows.

Vivek Hegde
maybeprascoder Kumar

Updates

maybeprascoder Kumar started this project — Oct 26, 2025 01:09 AM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.