FinNova 📈

A conversational financial document analyst powered by Amazon Nova and RAG — built in a few hours to learn retrieval-augmented generation from scratch.

---

Inspiration

I wanted a hands-on way to learn RAG in preparation for a larger upcoming project. The hackathon was the perfect excuse to spend a few hours going from zero to a working conversational AI over financial documents — and to experiment with Amazon Nova as the LLM backbone.

---

What It Does

FinNova is a Streamlit chat app that lets you ask questions about financial documents:

• Pre-indexed knowledge base — financial PDFs in docs/ are chunked, embedded, and stored in ChromaDB at startup
• Upload your own PDF — tax returns, financial statements, or any document you want the assistant to reference as personal context
• Conversational memory — follow-up questions work naturally; the assistant understands pronouns and references from earlier in the chat
• Two-LLM pipeline — a fast rephraser rewrites your question into a standalone query using chat history, then Nova answers using the retrieved documents

The retrieval scoring uses cosine similarity over embedded chunks. For a query $q$ and document chunk $d$, the relevance score is:

$$\text{score}(q, d) = \frac{\vec{q} \cdot \vec{d}}{|\vec{q}||\vec{d}|}$$

The top $k=5$ chunks are passed as context to the answering LLM.

---

How I Built It

Tech Stack

Layer	Technology
UI	Streamlit
Main LLM	Amazon Nova (nova-pro-v1) via langchain_amazon_nova
Rephraser LLM	Groq llama-3.3-70b-versatile (temperature=0)
RAG chains	langchain_classic — create_history_aware_retriever, create_retrieval_chain, create_stuff_documents_chain
Vector store	ChromaDB
Embeddings	OpenAI text-embedding-3-large

Learning Arc

1. Started with a basic prompt | llm chain and a simple retriever (the CLI in main.py)
2. Learned LangChain's conversational chain abstractions and refactored app.py to use create_history_aware_retriever + create_retrieval_chain for proper memory
3. Wired in a second LLM (Groq) as a zero-temperature rephraser to improve retrieval quality on follow-up questions

Key constraint: Used ChatAmazonNova directly rather than going through AWS Bedrock, due to credential complexity and request limits during the hackathon.

---

Accomplishments That I'm Proud Of

• Went from knowing nothing about RAG to shipping a working conversational RAG app with memory in a few hours
• The two-LLM architecture — using a cheap, fast model to rephrase and a powerful model to answer — is a pattern I'll carry forward
• Learned by building, not just reading docs

---

Challenges

• LangChain chain abstractions — create_history_aware_retriever, MessagesPlaceholder, and create_stuff_documents_chain all had subtle wiring requirements that took time to understand
• Two-LLM coordination — getting the rephraser output to feed cleanly into the retriever, then into the answering chain, required understanding how LangChain passes state between chain components
• @st.cache_resource and session state — cached LLM instances don't re-initialize when the temperature slider changes, a subtle Streamlit gotcha

---

What's Next for FinNova

The current version is intentionally simple — a foundation for learning. The next version is a full rethink:

• Multimodal input — voice dictation as the primary input method
• LaTeX output — financial formulas and equations rendered as $\LaTeX$ instead of plain text, e.g.
• Voice → LaTeX editor — the real goal: a tool that lets you speak financial calculations and get back structured, editable math notation
• Beyond RAG — richer document understanding using Nova's multimodal capabilities (tables, charts, scanned PDFs)

Built With

• amazon-nova
• chroma
• langchain
• openai
• python
• rag
• streamlit