Why Datasite Stopped Building Its Own Parsing - Adit Abraham, Reducto & Raj Bakhru, Datasite

Highlights

• Datasite handles approximately 60% of all transactions valued over $100 million.

• The platform manages over 500,000 monthly active users and billions of pages of unstructured data across 55,000 data rooms annually.

• Datasite's top-line revenue grew by roughly 40% over the eighteen months following the acquisition of Blue Flame.

• A custom benchmark called VDR Bench tests hundreds of questions against a standard virtual data room of hundreds of documents to evaluate retrieval and parsing performance.

• Parsing optimization through Reducto improved overall recall rates by 22% compared to baseline setups.

Timeline

1. 00:00:00 - 07:18:00

   Transitioning from Internal Parsing to External Tools

   The discussion covers the evolution of Blue Flame from an early vertical AI startup into Datasite's broader ecosystem. Operating virtual data rooms requires processing highly sensitive, unstructured financial data under strict security and regulatory frameworks. Maintaining internal parsers became unsustainable due to endless edge cases in financial spreadsheets and visual data. Shifting this burden to Reducto allowed Datasite's engineering organization to concentrate on building advanced AI agents and robust retrieval-augmented generation platforms.

   • Blue Flame originally built its own document ingestion and parsing pipeline to conserve capital as a bootstrapped startup.
   • Edge cases in financial documents, such as complex spreadsheet merge cells and mismatched chart legends, consumed excessive engineering support time.
   • Datasite acquired Blue Flame to integrate virtual data room assets and proprietary company data into an AI-driven workflow.
   • Partnering with Reducto eliminated the need to maintain custom parsing infrastructure and redirected engineers toward high-value application features like GraphRAG.
2. 07:18:00 - 16:00:00

   Scaling Engineering and Strategic Build vs. Buy Decisions

   The founders reflect on the trade-offs between remaining independent and joining an established enterprise player with massive proprietary data sets. Competing against frontier labs required immense capital and deep data access, making the Datasite acquisition a strategic advantage. Within the engineering department, leadership prioritizes directing developer bandwidth toward direct business growth and new product monetization rather than reinventing solved problems like document parsing.

   • Datasite evaluated raising hundreds of millions in venture capital before accepting an acquisition offer that provided immediate access to proprietary virtual data room assets.
   • The engineering organization comprises approximately 250 developers divided across five primary revenue-generating initiatives.
   • AI-assisted coding creates a barbell effect that increases the output of senior and junior engineers while reducing routine documentation burdens.
   • Outsourcing underlying foundational layers maximizes the efficiency of limited engineering talent.
3. 16:00:00 - 00:19:40

   Model Evaluation, Cost Optimization, and Architecture Q&A

   The closing segment details the technical stack and evaluation methodologies used to maintain high standards in financial data processing. Datasite utilizes VDR Bench to measure performance against frontier models and internal iterations. The findings confirm that robust document parsing and content extraction serve as the primary drivers of retrieval accuracy in complex financial workflows.

   • Finance users prioritize answer accuracy above all else, requiring continuous balancing of quality, speed, and token cost across different LLMs.
   • Datasite's ingestion architecture processes terabyte-scale document drops, routes content through Reducto, stores vectors in TurboPuffer, and executes GraphRAG queries using Python.
   • VDR Bench evaluates system performance by running hundreds of targeted questions against a standard test data room containing hundreds of files.
   • Accurate document extraction accounts for the majority of retrieval success, outperforming downstream adjustments in the answering layer.