How FinRAG Works

What is RAG?

Retrieval-Augmented Generation (RAG) is a technique that combines search with AI. Instead of relying solely on a model's training data, RAG retrieves relevant documents first, then uses them to inform responses. In financial research, this means searching across earnings calls, SEC filings, charts, and news -- using semantic understanding, not just keywords.

Traditional keyword search fails for finance: “revenue beat expectations” and “earnings topped estimates” mean the same thing but share almost no words. Embedding-based search captures this semantic similarity. Click each stage below to see how:

How Sentiment Analysis Works

FinBERT is a language model fine-tuned specifically for financial text. It classifies text as positive, negative, or neutral. But knowing the label isn't enough -- we need to understand why.

SHAP (SHapley Additive exPlanations) uses game theory to attribute the prediction to individual tokens. Each token gets a SHAP value showing how much it pushed the prediction toward or away from a given class. This is critical for trust in financial AI -- you need to know if the model is responding to meaningful signals or noise.

Why Multimodal?

Financial events produce data across multiple formats. An earnings report generates an audio recording (the call), a PDF filing (the 10-K), news articles (analyst reactions), and chart movements (price action). Analyzing only one modality means missing context from the others. FinRAG embeds all four using a single model, enabling cross-modal search.