Hebbia Tests Astra: Boosting Financial Research Efficiency with a Leap in PPT Generation

In the field of financial research, the value of AI models often comes down to two of the most practical steps: whether it can precisely locate the original sources that support decisions from vast amounts of documents, and whether it can quickly turn research conclusions into presentable, finished materials. After testing OpenAI's Astra model, the Hebbia team offered a fairly direct verdict—on the two tasks of citation source selection and PowerPoint generation, Astra showed double-digit performance improvements compared with the other models they tested.

Citation Source Selection: A Key Capability in the Data Room Scenario

Hebbia calls the first advantage "citation source selection," meaning that when faced with a question, whether the model can find the correct document or the correct source to answer it. This capability sounds basic, but in financial scenarios it directly determines whether AI output can be used.

The reason is that financial due diligence, investment research, and deal evaluation often involve data rooms containing thousands of documents. If the model finds the wrong file, even if the subsequent reasoning is fluent, the conclusion is left hanging. Astra's relative advantage in this step means that it is closer to the way professional researchers work when it comes to retrieval and locating original sources.

From Hebbia's description, this improvement is not a marginal one. They use "double-digit growth" to describe the lead over other tested models, indicating that the gap cannot be closed simply through parameter tuning or prompt engineering.

Data Room is a core work scenario in financial due diligence, investment research, and deal evaluation, usually referring to a controlled digital space that centrally stores thousands of legal, financial, and operational documents. In M&A transactions, the seller sets up a data room for the buyer and its advisory team to conduct due diligence. The number of documents often runs into the thousands, covering contracts, financial statements, intellectual property files, and more. Traditionally, analysts need to manually search and cross-check, which takes weeks or even months. For an AI model to play a role in this scenario, it must first solve the problem of "finding the right file"—that is, precisely locating the original source relevant to the question from massive unstructured documents. This is exactly why Hebbia emphasizes "citation source selection": financial decisions have extremely high requirements for traceability, and any conclusion must have a clear documentary basis. Incorrect model retrieval will cause all subsequent analysis to fail, so this capability is the first threshold for AI to enter financial workflows.

PowerPoint Generation: Turning Research into Decision-Ready Deliverables

The second advantage lies in PowerPoint generation. Hebbia particularly emphasizes that presentations in financial scenarios are not typesetting exercises; the views they carry need to be supported by solid research and ultimately influence capital deployment decisions worth tens of millions or even hundreds of millions of dollars.

What surprised Hebbia about Astra was that it can produce highly precise outputs and deliverables and follow user-defined specifications. Artifacts such as presentations have implicit requirements for format, structure, and content organization. The model must simultaneously balance factual accuracy and presentation standards, which is far more difficult than generating a passage of free text. Hebbia's exact words were that Astra's leap in this direction is "truly astonishing compared with other models."

This also explains why Hebbia discusses the two capabilities together: source selection ensures the traceability of the views in the PPT, while document generation capability determines whether research results can be efficiently implemented. Only when the two are combined does AI truly enter the workflow of financial research, rather than remaining at the stage of a Q&A toy.

Artifact in the context of AI-assisted professional work refers to structured outputs generated by the model that can be directly used for business delivery, such as presentations, reports, and spreadsheets. Unlike free text, artifacts need to follow clear formatting specifications, structural logic, and content organization requirements. In financial scenarios, PowerPoint presentations are the main form for presenting investment research conclusions, transaction recommendations, and due diligence findings, and their quality directly affects the decisions of investment committees and clients. Generating qualified financial PPTs requires not only fluent language but also factual accuracy, traceable data, layouts that conform to institutional templates, and views supported by solid research. The model needs to handle both content generation and formatting constraints at the same time, which is far more difficult than open-domain question answering. Hebbia places PowerPoint generation alongside citation source selection precisely because together they form the closed loop of financial AI workflows: the former ensures that views are traceable, and the latter ensures that results are deliverable.

Implications for Financial AI Applications

Hebbia's feedback points to a broader trend: evaluating the value of AI models in professional scenarios is shifting from general benchmarks to the quality of specific deliverables. Whether citations are correct and whether PPTs conform to user specifications—these indicators are closer to the reasons enterprises are willing to pay than abstract reasoning scores.

For enterprises currently selecting financial research tools, the double-digit improvements Astra showed in Hebbia's tests are worth noting, but they should also validate them against their own data room scale, document types, and delivery specifications. Hebbia's conclusions come from its own business scenarios, and experience sharing from a single source cannot replace systematic evaluation.

At least judging from this feedback, Astra has crossed a practical threshold in high-precision document location and structured deliverable generation. For fields such as financial research, where the tolerance for error is low, this matters far more than whether the model can chat.