Argusa AI Challenge 2025 · Runner-Up 🥈

Secured 2nd place at the Argusa AI Challenge 2025 by building ARGRAG, a RAG system for complex enterprise document corpora with multi-modal analysis, metadata intelligence, and real-time performance.

Competed in the Argusa AI Challenge 2025 alongside Eisha Tir Raazia and secured 2nd place out of 20 teams with our project ARGRAG.

The Challenge

The hackathon focused on a common enterprise pain point: extracting reliable, explainable business insights from large volumes of scattered, unstructured internal documents. We had to:

  • Ingest heterogeneous document collections across many formats
  • Answer arbitrary business questions with high accuracy
  • Provide explainability via traceable sources and supporting evidence
  • Deliver interactive performance suitable for real users

Our Solution: ARGRAG

We built ARGRAG, a retrieval augmented generation system designed for complex enterprise document corpora.

Intelligent Document Processing

  • Support for 15+ file formats including PDFs (with OCR), Office documents, images, archives, and source code
  • Robust extraction and normalization across inconsistent inputs
  • Automated handling of archives (ZIP, TAR, 7Z, etc.)

Multi-Modal Analysis

  • Vision AI to interpret images, diagrams, charts, and embedded visual evidence
  • Automatic image description generation for context enrichment

Metadata Intelligence

  • LLM-generated corpus understanding to identify projects/teams, detect technology stacks, extract key concepts, and map document relationships

Multi-LLM Query and Answer Pipeline

  • Query augmentation using corpus metadata to expand user questions with relevant context
  • Vector search over ChromaDB for high-recall retrieval
  • Answer generation with explicit source attribution for explainability

Incremental Updates

  • Smart file watching with checksum-based caching
  • Efficient per-file incremental updates without full reindexing
  • Real-time change detection (1s polling)

Performance

  • Subsequent boot: < 5 seconds
  • Query response: 2 to 5 seconds
  • UI load time: < 1 second
  • File change detection: real-time (1s poll)
  • Incremental update: per-file basis

Tech Stack

  • Backend: FastAPI (Python 3.11+)
  • Vector DB: ChromaDB
  • Relational DB: PostgreSQL (document metadata and checksums)
  • Embeddings: SentenceTransformers (local, free)
  • Vision AI: Moondream and LLaVA (can run locally)
  • LLM: GPT-OSS (120B) via Ollama Cloud (large context, free within limits)
  • Frontend: React + Material UI

Team