StreamingRAG Real-Time Anomaly Detection
Overview
StreamingRAG is framed as a single end-to-end system: real-time anomaly detection coupled with a built-in RAG layer for grounded interpretation. In practice, the SRAG-V2 pipeline concepts (stream ingestion, frame scheduling, inference routing, temporal context updates, and knowledge-layer feedback) are treated as part of the StreamingRAG implementation direction and map to the streaming-oriented modules.
At the current milestone level, the strongest demonstrated outcome was detecting a car crash in sample video input while preserving the retrieval-grounded response flow.
Source code repositories:
Project Features and Team Work
The project includes the following core capabilities:
- A FastAPI service that serves RAG-based responses and a lightweight web interface for interacting with the system.
- PostgreSQL + pgvector retrieval that enables semantic similarity search over ingested knowledge.
- Dual response behavior (grounded vs. general) with response metadata such as
modeandmax_similarityto indicate retrieval confidence. - Prompt-template routing for grounded and general answer strategies through dedicated prompt files.
- Local model inference via Ollama, enabling privacy-friendly development without reliance on hosted LLM APIs.
- A modular structure with components for ingestion, persistence, pipeline orchestration, and streaming/knowledge workflows.
- An anomaly-analysis pipeline that includes inference, knowledge graph updates, temporal event-state identification, and adaptive question generation, with SRAG-V2 modules serving as the concrete architecture reference integrated into the streaming-side design.
Project Outcomes
StreamingRAG demonstrates a practical combined architecture rather than two disconnected efforts. It integrates computer-vision signals, temporal reasoning, and retrieval-grounded generation into one iterative pipeline, with car-crash detection in sample footage as a concrete progress milestone.
