FinSage: Agentic RAG Financial Advisor

An AI-powered financial Q&A chatbot that thinks before it answers. A Planner decides on every turn whether to retrieve knowledge from a document base or lean on conversation context, with no unnecessary lookups and no hallucinated facts.

---

The Problem This Solves

Most chatbots fall into one of two traps:

Trap	What goes wrong
Always retrieve	Slow, pulls irrelevant docs for simple follow-ups
Never retrieve	Hallucinates specific rates, rules, and figures

Financial advice has zero tolerance for made-up numbers. FinSage adds a routing layer between the user and the LLM that asks "does this question need grounded knowledge?" before every response. If yes → retrieve and answer with context. If no → answer from the conversation alone, faster and cleaner.

---

Architecture

---

What You Learn Building This

• How RAG (Retrieval-Augmented Generation) works end to end
• How to build a FAISS vector index with sentence-transformers
• How agentic routing reduces unnecessary LLM calls and cost
• How to structure a modular AI pipeline where every layer is swappable
• How to manage conversation history properly in a stateless way
• How to wire an LLM backend to a Gradio chat UI
• Basics of prompt engineering: system prompts, context injection, and history formatting

---

Project Structure

finsage/
├── agentic_rag_financial_advisor/
│   ├── planner.py       # Routes query → rag or chat_history mode
│   ├── retriever.py     # FAISS vector search with sentence-transformers
│   ├── rag_pipeline.py  # Prompt assembly + LLM call
│   ├── llm.py           # LangChain ChatOpenAI wrapper
│   ├── config.py        # API key loader from env vars
│   └── utils.py         # Chat history cleaner
├── app.py               # Gradio entry point
├── requirements.txt
└── README.md

---

Quick Start

git clone https://github.com/c-kiplimo/FinSage.git
cd FinSage

python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate

pip install -r requirements.txt

export OPENAI_API_KEY=sk-...     # required
# export ANTHROPIC_API_KEY=...   # optional
# export GEMINI_API_KEY=...      # optional

python app.py                    # opens http://localhost:7860