Building a high-accuracy financial document intelligence system with Gemini 2.5 Pro

Pascal AI is an AI-native operating system designed for investment funds, operating at an institutional scale to process millions of pages of filings, memos, and models. Their mission is to turn a firm’s internal and external data into an AI-powered advantage, helping analysts and CIOs make faster, data-driven decisions.

To build the knowledge graph that powers their agentic workflows, Pascal AI needed a document intelligence system capable of converting diverse, complex financial documents into structured text with exceptional accuracy.

Financial documents present unique, stubborn challenges for programmatic parsing. Prior to integrating the Gemini API, the Pascal AI team tested various OCR tools and large language models, encountering persistent technical hurdles:

• Complex visual data: Extracting accurate data from multi-axis charts and trend graphs is non-trivial. Other models frequently hallucinated values not present in the original visuals, creating unacceptable reliability issues.
• Intricate table structures: Financial statements often feature merged cells and span multiple pages both horizontally and vertically. Standard extraction libraries often failed to preserve this structure, losing critical context such as currency units or time periods.
• Varied document quality: Data sources range from digital-native filings to low-resolution, scanned PDFs, making rigid parsing logic brittle.

Pascal AI required a parsing layer that could handle this complexity without hallucination.

To overcome these challenges, Pascal AI integrated Gemini 2.5 Pro via LangChain as the core of their document intelligence stack.

According to Kanav Anand, AI Lead at Pascal AI, the model’s multimodal reasoning significantly boosted accuracy. Unlike previous solutions, Gemini 2.5 Pro minimizes hallucinations and accurately transforms complex graphs and charts into structured markdown tables, preserving vital financial context.

To measure success, Pascal AI utilizes an internal evaluation set, tracking edit distance error-rate to determine how close the parsed output is to the original text. Gemini 2.5 Pro achieved a low 4% edit distance error rate, performing 2x more accurately than the next best model tested. Furthermore, the model achieved 100% element-wise accuracy, correctly identifying structural components like tables, paragraphs, and headers.

Beyond raw accuracy, the Gemini API improved development velocity. By solving complex document intelligence problems primarily through prompt engineering rather than brittle custom logic, the team can iterate quickly to support new document types as they become available.

Looking ahead, Pascal AI aims to push towards near 100% parsing accuracy by experimenting with advanced methods, including model orchestration and fine-tuning for domain-specific financial reporting.

To start building with Gemini models, read our API documentation.