Semicon-Insight: Intelligent Financial RAG Analyst 📈

An Enterprise level mulitomodal RAG system, designed for Semiconductor industry(NVIDIA VS AMD) 10-k financial report deep analysis

📖 Background

Traditional rag system would face two difficulties when dealing with enterprise 10k financial report.

1. Complex tables analysis losses efficacy Cross-page tables in pdf might be interpreted as garbled characters, preventing LLM from answering specific financial numerical questions.
2. Lack macroscopic comparison ability Simple vector search cannot solve "Compare the two company's strategies" such macropic problems. ** Semicon-Insight**: By introducing Multimodal-Analysis, Hybrid-Chuncking and Router-Chucking, I realize efficient unstructured text and structured financial data.

System Architecture

The system utilized Router-based Agentic RAG structure, pick up search strategy based on user intention.

graph TD
    User[User Query] --> Router{Router Query Engine}
    
    subgraph "Data Ingestion Layer"
        PDF[10-K PDFs] -->|LlamaParse| MD[Markdown]
        MD -->|MarkdownElementNodeParser| Nodes[Text & Table Nodes]
    end

    subgraph "Indexing Layer"
        Nodes -->|Embedding| Qdrant[(Unified Qdrant Vector DB)]
        Nodes -->|Summary| SumNVDA[NVIDIA Summary Index]
        Nodes -->|Summary| SumAMD[AMD Summary Index]
    end

    Router -- "Specific Fact / Comparison" --> ToolA[Vector Search Tool]
    Router -- "NVIDIA Overview" --> ToolB[NVDA Summary Tool]
    Router -- "AMD Overview" --> ToolC[AMD Summary Tool]

    ToolA -->|Hybrid Search + Re-ranking| LLM
    ToolB -->|Tree Summarize| LLM
    ToolC -->|Tree Summarize| LLM

    LLM --> Response

Loading