FinAssist-AI: A Fully Offline Intelligent Financial Document Analysis System

FinAssist-AI is a full-stack financial document analysis application based on the RAG architecture. It supports local deployment of the DeepSeek-R1 inference model to enable intelligent Q&A and analysis of financial data without an internet connection. This project addresses issues such as data leakage risks, network dependency, and high API costs associated with traditional cloud-based AI solutions, ensuring data privacy and being suitable for various financial scenarios.

Financial data analysis has extremely high requirements for accuracy and privacy. Traditional cloud-based AI solutions have data leakage risks, network dependency, and high API call costs. Especially when processing sensitive financial statements and contract documents, institutions are cautious about uploading them to third-party cloud services. FinAssist-AI adopts a fully offline architecture, allowing users to complete the entire process from document parsing to intelligent Q&A locally, ensuring data privacy and reducing network dependency.

FinAssist-AI adopts a modern full-stack architecture: The frontend is based on Next.js 16 (with Turbopack enabled), providing modes such as server-side rendering and static generation. Turbopack replaces Webpack to improve the speed of hot reloading during development. The backend uses the FastAPI framework, which is based on Starlette and Pydantic, ensuring API type safety and performance, and is suitable for handling LLM streaming responses.

Document Parsing Layer: Docling Core

Financial documents have complex layouts (tables, charts, multi-column text), which ordinary tools struggle to recognize. Docling Core can identify semantic structures, convert tables into structured data, retain paragraph hierarchies, and provide a foundation for high-quality text chunking and vectorization for RAG retrieval.

Vector Storage: Local ChromaDB

It uses the lightweight embedded ChromaDB to store document embedding vectors, requiring no additional service deployment. Data is saved in the local file system, and queries generate no network traffic. It supports multiple measurement methods such as cosine similarity and Euclidean distance.

Inference Engine: Local Deployment of DeepSeek-R1

It supports local operation of the open-source DeepSeek-R1 inference model (which excels in mathematical reasoning and code generation), enabling high-quality inference capabilities without an internet connection. It implements a streaming response mechanism, with real-time output display to enhance the interactive experience.

Retrieval-Augmented Generation (RAG) is the core mechanism, divided into four stages:

1. Document Ingestion: Users upload financial documents such as PDFs. Docling Core parses the layout, extracts structured text, and retains semantic information such as chapter hierarchies and table structures.
2. Text Chunking and Vectorization: Text is split into chunks along semantic boundaries, converted into high-dimensional vectors by the embedding model, and stored in ChromaDB to build an index.
3. Retrieval: User queries are converted into vectors, and ChromaDB searches for the most similar text chunks to achieve semantic retrieval (regardless of whether keywords are exactly the same).
4. Generation: The retrieved relevant text chunks are used as context and submitted to DeepSeek-R1 along with the query to generate accurate and traceable answers.

FinAssist-AI is suitable for various scenarios:

• Investment Analysis: Quickly extract key indicators from financial reports and compare quarterly performance.
• Audit: Check the consistency of contract terms and identify risk points.
• Small and Medium Financial Institutions: No need for expensive cloud services, solves data compliance issues, and local servers support daily needs.
• Education: Financial major students analyze real financial reports and learn to extract information.

• Hardware Configuration: Choose the appropriate model size based on different parameter versions of DeepSeek-R1, considering memory and GPU resources.
• Development Environment: Docker configurations are provided to simplify dependency installation.
• Production Environment: Configure sufficient memory and GPU to ensure inference speed.
• Document Processing: For large-scale processing, it is recommended to implement an asynchronous queue to avoid blocking the interface.

FinAssist-AI represents an important direction for financial AI applications: providing intelligent analysis while ensuring data privacy. With the performance improvement of open-source large models and the maturity of local deployment tools, similar offline solutions will become more popular. For fintech developers, it is an excellent example for learning RAG architecture and local LLM deployment, and its code and architecture are worth in-depth study.