When we built Parsewise, we wanted users to trust its answers. Data rooms rarely contain a single clean source. They usually hold documents in several formats, and the same figures may conflict or be revised over time. Our OfficeQA article covered what it takes to answer a difficult question from this kind of material.
In addition, an analyst's work rarely ends with the first answer. They need to inspect the underlying data, follow unexpected results, and ask new questions as they go.
This article looks at what allows Navi to do that across very large document collections while keeping its answers accurate and traceable.
The foundations
Scale starts before Navi receives a question. Parsewise first turns the documents into evidence that Navi can query and reuse.
- Parsewise extracts each document and stores its contents in a structured form. Navi can return to this evidence when a new question arises without processing the source document again.
- When an investigation requires a view of the full corpus and results that can be saved, business users or Navi can create Agents to perform exhaustive search grounded in the document evidence.
Navi at scale
These foundations allow Navi to answer difficult questions across very large document collections without loading every page into the model at once.
Navi starts by querying the structured data to narrow the investigation to the evidence relevant to the current question. It brings only the records and source material needed for the next step into the model, rather than the entire collection. From there, Navi can inspect the candidate values, inconsistencies, reasoning, and source pages behind an Agent's answer. This allows it to compare revisions, check whether values refer to the same period or definition, and explain why a particular source was selected.
Navi also has an isolated sandbox for more complex analysis over approved project data. Within it, Navi can combine evidence, perform calculations, and explore relationships in the data without changing the underlying information.
Navi can follow up on results, work through several analytical steps, and reach evidence-backed conclusions without requiring the user to prepare a new workflow or export the data elsewhere.
Conclusion
Business users can start with a broad question even when the relevant evidence may be spread across thousands of pages. Navi narrows the collection, compares sources, and carries the analysis forward as new questions emerge, incrementally building a data foundation.
Our previous article covered the latest additions to Navi, and there is more to come. We're building new tools to help Navi take on larger and more complex investigations, alongside improvements to how users work with it.