Datasite explains why it switched from DIY parsing to Reducto

AI Engineer brought together Reducto co-founder Adit Abraham and Datasite global chief technology officer Raj Bakhru to detail why Datasite stopped building its own document parsing. Bakhru said the virtual data room provider, which he said handles about 60% of transactions over $100 million and runs 55,000 data rooms a year holding billions of pages for more than 500,000 monthly active users, moved ingestion to Reducto after its scrappy in-house pipeline collapsed under edge cases.

The detail that broke it was not a missing model. Bakhru said a private equity firm flagged a wrong answer where the shade of blue on a pie chart legend did not map correctly for the vision model. Pass the same data as a bar chart and the model got it right. Finance spreadsheets were worse, he said, with merged cells, XLSB and macro-enabled and template files that a human struggles to read, let alone a parser.

That is why the mechanism now matters more than the model. Bakhru said Reducto runs inside Datasite's environment against content that has never left its Azure platform before, because the rooms hold material non-public information and must satisfy ISO and other regulatory regimes. The ingestion pipeline takes volume that can spike by a terabyte on a bank drop, passes it through Reducto, lands vectors in Turbopuffer and then feeds a proprietary Python GraphRAG stack that answers questions inside the room.

Bakhru framed it as a build versus buy call for 250 engineers. He said Datasite has grown about 40% top line since the BlueFlame acquisition a year and a half ago, when the 50-person startup was bought over dinner after pitching a $5 million check, and now must generate roughly $200 million in incremental revenue a year to hit a 20% growth ambition on a business around $1 billion in revenue. He said he wants those engineers split across five revenue-generating initiatives, not support cases for extraction that a vendor solves daily, and that bringing proprietary data closer to models with Grata's 20 million private company records and the VDR corpus already saves token cost and runtime.

Friction has not disappeared. Bakhru said the team still benchmarks models constantly for cost, speed and quality, with quality non-negotiable for finance users. He cited tool-call spamming differences between Sonnet variants and Opus, and said Datasite runs VDR Bench, an internal eval of hundreds of documents and questions, on every change and expects to beat dropping the same files into frontier labs' projects. He said recall improved 22% because of Reducto's parsing, not the answering layer, and that most work is on extraction. VDR Bench is not open sourced, he said, though a paper with OpenAI is under consideration.

Demand for that extraction keeps rising beyond Q&A toward agents that run diligence and draft decks. Even outside this talk, Datasite said Oct. 29 it is partnering with Prudentia Sciences to link approved transaction materials in data rooms to life sciences diligence, per GlobeNewswire via Manila Times. The practical consequence Bakhru described is simple. If parsing fails, GraphRAG fails, and no agent can paper over it.