Openqa: Building a Reasoning Learning Model for the Financial Domain from Scratch, Integrating Tool Calling and RAG

This article introduces the open-source project Openqa, which builds a reasoning learning model for the financial domain from scratch, integrating tool calling and retrieval-augmented generation (RAG) capabilities. It aims to address the pain points of applying general large language models in the financial domain and explore the path to building domain-specific AI. The project is maintained by Pablo Garcia Amolina, with source code hosted on GitHub (link: https://github.com/pablogarciaamolina/Openqa), and was released on June 9, 2026.

General large language models (such as GPT-4, Claude) face four major challenges in financial applications:

1. Insufficient domain knowledge: Lack of professional in-depth knowledge, prone to hallucinations or outdated information;
2. Limited reasoning ability: Financial analysis requires complex calculations and multi-step decision-making, and general models are not reliable enough;
3. Real-time data acquisition: Financial markets change rapidly and need real-time data support;
4. Tool integration needs: Need to call external tools such as calculators and databases. The Openqa project is designed to address these challenges, integrating three core capabilities: reasoning learning, tool calling, and RAG.

Openqa's architecture integrates three key technologies:

1. Reasoning Learning Model

Train/fine-tune a domain-specific model from scratch, supporting domain-customized training (financial corpus), reasoning ability enhancement (chain-of-thought training, process supervision), and high controllability and interpretability.

2. Tool Usage Capability

Implement a Function Calling mechanism that can call calculation tools (Python interpreter), data query tools (financial databases/APIs), and analysis tools (professional libraries) to compensate for model limitations.

3. Retrieval-Augmented Generation (RAG)

Build a financial knowledge base (regulations, reports, news, etc.), dynamically retrieve relevant knowledge and inject it into the context to solve the problems of knowledge cutoff and hallucinations.

Openqa's application scenarios in the financial domain include:

• Investment research and analysis: Automatically retrieve financial reports/data, tools ensure calculation accuracy, and RAG ensures information reliability;
• Risk assessment reports: Multi-step reasoning to analyze risk exposure, call models to calculate indicators such as VaR/CVaR;
• Compliance checks: Retrieve regulatory provisions to determine business compliance;
• Customer consultation services: Understand questions, retrieve product information, calculate returns, and provide personalized suggestions;
• Market intelligence summaries: Monitor news, generate dynamic summaries and trend analysis.

Building Openqa requires solving the following technical challenges:

1. Model training infrastructure: Adopt PEFT technology (LoRA/QLoRA) to reduce training costs;
2. Tool calling reliability: Design robust protocols and retry mechanisms;
3. RAG retrieval quality: Hybrid retrieval strategy (vector + keyword + knowledge graph);
4. Multi-module coordination: Clear interfaces and orchestration logic to ensure information flow;
5. Balance between latency and cost: Find a balance between model size, accuracy, and response speed.

Dimension	Openqa Domain-Specific Solution	General Large Model Solution
Domain Knowledge	Deeply customized, accurate and reliable	Generalized knowledge, may be outdated
Reasoning Ability	Optimized for domain tasks	General reasoning, may be insufficient for professional tasks
Real-time Data	Obtained via tools/RAG	Dependent on model training data cutoff
Calculation Accuracy	Ensured by tools	May have calculation errors
Development Cost	Requires domain data and customized development	Out-of-the-box, API calls
Controllability	High, interpretable	Low, black box
Domain-specific solutions are suitable for enterprise-level applications (high requirements for accuracy and controllability), while general solutions are suitable for rapid prototyping and general tasks.

Community value of Openqa as an open-source project:

• Reference for domain AI construction: Demonstrates the path to building domain-specific AI from scratch, providing architectural references for fields such as healthcare and law;
• Example of technology integration: The integration model of reasoning learning + tool calling + RAG provides specific implementation references;
• Foundation for financial AI: Reduces the threshold for financial AI application development, providing basic components and best practices.

Openqa represents the trend of AI evolving from general to domain-specific, proving that integrating reasoning learning, tool calling, and RAG can build professional systems with both flexibility and accuracy. For the financial industry, AI systems that are controllable, interpretable, and traceable are more valuable than black-box general models. As model efficiency improves and the tool ecosystem matures, domain-specific AI solutions like Openqa will be widely applied in various industries.