This is a submission for the Redis AI Challenge: Real-Time AI Innovators.
What We Built
Ever spent too long digging through a folder of documents, trying to remember where you saw that one useful paragraph? We built a tool to make that a lot easier.
This project is an AI-powered semantic document search application that supports PDF, TXT, and CSV uploads. Once a document is uploaded, it’s processed using a Sentence Transformer model to extract embeddings that capture the meaning of the content and not just the words.
These embeddings, along with metadata like filenames and content snippets, are stored in Redis Cloud as FLOAT32 vectors. Redis 8’s HNSW-based vector search index enables fast and accurate nearest neighbor searches. Users can ask questions in natural language, and the system returns the top three most relevant documents, complete with the matched content and a similarity score.
What makes it practical is its real-time behavior. When users update or delete documents, the associated embeddings and metadata are immediately updated or removed from Redis. That way, search results stay clean and current. On the backend, MongoDB stores persistent user and file metadata, complementing Redis’s real-time data handling with durable storage.
Key Features
- Multi-format support – Works with PDFs, TXT and CSV.
- Real-time search – New uploads and deletions instantly reflected in results.
- Semantic understanding – Uses transformer embeddings to find meaning, not just keywords.
- Accurate retrieval – Redis 8 HNSW vector search delivers precise matches in milliseconds.
- Metadata-rich results – Returns filename, snippet, and similarity score for context.
- Hybrid storage – Redis for speed, MongoDB for long-term persistence.
Demo
Backend URL - Render
Frontend URL - Vercel
Demo Video
How We Used Redis 8
Redis Cloud powers the real-time vector similarity search in this project. For each document uploaded, embeddings are generated using a transformer model and stored as FLOAT32 vectors inside Redis. Redis 8’s HNSW indexing allows for efficient and accurate semantic queries with low latency.
But it's not just about vectors. Each Redis hash also stores relevant metadata and content chunks, so the system can return not just which document matched, but exactly what part of the document did.
Updates and deletions are handled via Redis commands, keeping the vector index tightly in sync with user actions. It’s quick, clean, and doesn’t require any batch processing or manual cleanup.
This project demonstrates Redis’s strengths as a multi-model, real-time data platform that supports fast AI-driven workflows. MongoDB complements it by providing persistent storage for user information and document-level metadata, giving the system both speed and long-term reliability.
Top comments (0)