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

## Executive summary

Datasite's CTO, Raj Bakhru, details how his company integrated advanced AI into its platform, which handles 60% of all transactions over $100 million. The core challenge was parsing massive amounts of highly sensitive, unstructured data (VDRs). Bakhru emphasizes that the complexity of 'edge cases' (e.g., merged cells, chart interpretation in Excel) made building the parsing layer internally impractical. They strategically leveraged Reducto's tooling to solve the extraction problem, allowing their 250 engineers to focus on high-value, revenue-generating initiatives like building Graph RAG and advanced agents, rather than 'rebuilding the wheel' on data ingestion.

## Key takeaways

- The Scale of Unstructured Data: Datasite processes billions of pages of content across 55,000 data rooms annually, making the management of Material Non-Public Information (MNPI) and unstructured data a massive AI priority.
- The 'Edge Case' Problem in Parsing: Parsing financial documents is defined by 'all edge cases.' Issues like interpreting chart data (e.g., shade of blue on a legend) or handling complex Excel formats (like XLSB or merged cells) require specialized tooling, making internal development costly and time-consuming.
- Strategic Build vs. Buy Decisions: To maximize the impact of their 250 engineers, Datasite shifted focus from building foundational data extraction layers to building value-add, revenue-generating AI products (e.g., agents, Graph RAG).

## Technical details

- Data Ingestion and Architecture: The system uses a specialized ingestion pipeline that passes content through Reducto's tooling, which then feeds the data into a vector database (Turbo Buffer). The final layer is a Graph RAG platform, with queries executed using proprietary Python code.
- Performance Benchmarking (VDR Bench): Datasite uses a custom evaluation, 'VDR Bench,' which tests the system against a common virtual data room of hundreds of documents. They reported a 22% improvement in recall rate directly attributed to the superior content extraction capabilities of Reducto's parsing.
- AI Development Strategy: The focus for internal engineering talent is on solving the hardest business problems (e.g., building agents for complex diligence analysis) rather than optimizing foundational data layers. Optimization efforts are constantly balancing three factors: cost, speed, and quality.

## Practical implications

- When dealing with highly complex, proprietary data (like financial VDRs), outsourcing the foundational data extraction and parsing layer to specialized vendors can be a critical architectural decision.
- Internal engineering teams should prioritize building high-level, revenue-generating applications (agents, specialized RAG) rather than spending time on foundational support cases or optimizing solved data pipelines.
- The 'Build vs. Buy' decision must be weighed against the cost and time required to solve niche, high-complexity edge cases inherent in the data source.

## Topics

AI Engineering, Retrieval Augmented Generation (RAG), Unstructured Data Parsing, Build vs. Buy Strategy, Financial Technology (FinTech), Adit Abraham (Reducto), Raj Bakhru (Datasite)

Source: https://www.youtube.com/watch?v=0PqGT2L6cZM