Skip to content

RAG Architecture

Planned Architecture (Future Phases)

The RAG system is implemented in Phase 3 (Weeks 8–11). This document describes the planned design.


Overview

FinSight uses a hybrid retrieval strategy combining dense vector search (pgvector) and sparse keyword search (Elasticsearch) to retrieve the most relevant financial document chunks before passing them to an LLM.


Architecture Diagram

flowchart TD
    Q[User Query] --> E[Query Encoder]
    E --> VS[pgvector — Dense Search]
    E --> ES[Elasticsearch — Sparse Search]
    VS --> R[Reciprocal Rank Fusion]
    ES --> R
    R -->|Top-k chunks| P[Prompt Builder]
    P --> L[LLM — GPT / Claude]
    L --> A[AI Response]

Retrieval Strategy

Dense Retrieval (pgvector)

Sparse Retrieval (Elasticsearch)

Reciprocal Rank Fusion


Prompt Design


Context Window Management


Observability