Why LLMs fail on financial documents and how AI extraction can finally fix it

How does it work? Inside Desia’s VLM-powered parsing pipeline

‍By combining VLM-powered extraction, contextual enrichment, and spreadsheet logic preservation, Desia delivers end-to-end parsing built for the real workflows of investment professionals. It accelerates due diligence, reduces manual review, and gives teams confidence in the integrity of AI-driven analysis.

‍Smart file understanding

‍Desia automatically classifies and routes each file type (PDF, Word, Excel, image, and text) through its optimal processing path, ensuring every document is handled based on its actual internal structure and content, not just the file extension.

‍Conversion

‍Documents are normalized into standardized intermediate representations. Office documents are converted into PDFs in isolated, resource-constrained environments to preserve layout and formatting:

‍

• PDFs go through text extraction first, followed by page-to-image conversion.
• Excel spreadsheets are handled through a dedicated parsing path that preserves formulas, data types, and sheet relationships.
• Images are routed directly into the vision-language processing stage.
• Plain text files are segmented into chunks based on content structure and token limits.

This approach provides uniformity for downstream processing while preserving the structural and semantic integrity of each file type.

‍

‍Parsing

‍Documents are processed through an orchestration layer that coordinates multiple vision-language models. Each page is represented both as an image and as text, enabling models to capture layout, diagrams, tables, and written content in parallel. Processing begins in batches for efficiency and automatically falls back to page-level retries under rate limits or when quality thresholds are not met. Retry strategies apply progressive backoff and alternative processing paths to maintain throughput and reliability.
‍

‍
‍Hybrid extraction strategy

‍Extraction runs in parallel, combining visual and textual elements:

‍

• Visual extraction renders pages as images, preserving layout, formatting, tables, charts, and diagrams.
• Textual extraction provides the core content, preserves reading order, and exposes metadata.

Both representations are merged into a single input, allowing the model to cross-reference them. This enables richer outputs, including descriptive analysis of charts, diagrams, and other visual elements. This allows Desia to interpret not just written content but also the charts, tables, and figures that are central to financial analysis. Crucially, Desia parsing solution also handles scanned documents, where traditional AI often fails.
‍

‍
‍Spreadsheet specialization

‍Desia is built to handle various spreadsheets with precision. The files are processed using a dedicated pathway that maintains their inherent structure. Instead of flattening spreadsheets into PDFs, the parser preserves:

‍

• Cell coordinates and relationships
• Formula logic and dependencies
• Data types (dates, currencies, numerical precision)
• Advanced features such as pivot tables, charts, and conditional formatting

By preserving the logic behind the numbers, through ensuring accurate interpretation of spreadsheet semantics, Desia makes AI-powered due diligence and portfolio value creation more reliable and effective.

‍Context enrichment

‍After parsing, every document passes through a context enrichment pipeline designed to generate document-level intelligence. This stage is a crucial step that ensures individual page outputs are interpreted in relation to the full document. The enrichment process uses a document-level cache to persist embeddings and analysis results across the entire file. The full document is first stored in cache, enabling global context to be referenced throughout subsequent steps. Pages are then processed in batches, where each page-level analysis incorporates both the page content and references to the global cached context. This design enables:

‍

• Page summaries enriched with cross-document awareness
• Identification of relationships between sections, tables, and figures across different pages
• Detection of recurring themes or references that span multiple chapters or sections
• Consistency in terminology and interpretation across the entire document

‍

‍Validation

‍Extracted outputs undergo multi-level quality assurance. Visual and text-based results are cross-validated for consistency. Duplicate detection operates at both page and context levels, filtering out repeated phrases and sentences. Documents must meet configurable quality thresholds for completeness, coherence, and structural integrity. Failures trigger retries with adjusted strategies until acceptable results are achieved. The result is a cohesive, markdown-format, document-aware output where each page is not only parsed locally but also contextualized against the larger structure and narrative of the document.