Spotting Gaps Before They Matter: How Samaya Enhances Answer Accuracy in Real Time

Roberto Dessi Roberto Dessi

Why Do Gaps in Answers Matter?

Imagine you are a financial analyst preparing a quarterly market report. You ask Samaya a question like:

What signals are companies giving about labor markets amid an uptick in unemployment — are they freezing hiring, cutting headcount, or holding steady? Give me numbers and summarize your answer in a table.

Samaya gives you a clear and well-structured response within seconds. It cites reliable sources, presents data in a neat format, and follows your instructions precisely. Samaya is built to handle such knowledge-intensive queries with a high degree of accuracy.

Still, for particularly broad or multi-faceted questions, some specific aspects of information may not initially be included — not due to system failure, but because retrieving, synthesizing and reasoning over large volumes of information across many sources on the fly is extremely challenging, and can occasionally lead to certain details being left out. In practice, even small omissions may introduce friction. Users may need to cross-check additional sources to validate an answer, especially when preparing for high-stakes meetings. While Samaya’s Q&A and Agent systems already save our users a huge amount of time, we wanted to go one step further.

We asked ourselves: What if the system could identify potential information gaps in the summary before the user even notices? What if it could proactively retrieve the missing pieces seamlessly, without interrupting the flow of the interaction?

That’s why we built Samaya’s new self-correction capability into Samaya’s core engine. It runs quietly in the background, checking for data coverage and summary quality, and automatically pulling in additional evidences when needed. This way, users get even more comprehensive and trustworthy outputs, without any additional effort or delay.

How It Works

Samaya’s Q&A system is designed to retrieve relevant documents and generate a grounded, accurate summary within seconds. To maintain speed and avoid any added latency, our self-correction engine operates entirely in parallel with the main summary streaming system. It functions as a silent yet reliable quality assurance layer, constantly working to provide the most complete and accurate answers possible.

The following animation demonstrates the Q&A experience with self-correction engine enabled:

How Samaya’s Q&A system corrects its summary in real-time.

As shown in the diagram below, the whole system works in the following steps:

  • Quality checking for omissions: As Samaya’s Q&A system streams the summary, the self-correction engine simultaneously analyzes the query and the response in real time, evaluating whether any key information is missing or needs correction.
  • Launching targeted background searches: If any gaps are detected, the system launches precise, on-the-fly searches aimed at retrieving just the right amount of information to fill the gaps.
  • Correcting summary with newly fetched evidences: As new information is found and verified, the system updates the summary in real time with newly fetched evidences, ensuring a high-quality response.
Diagram of the self-correction engine
How Samaya’s self-correction engine works with real-time summary streaming.

Evaluation Results

We evaluate Samaya’s self-correction engine across two key dimensions:

Answer Coverage: We built a benchmark consisting of numerically intensive queries that require reasoning and aggregating information across many data sources and presenting the results in tabular form. To assess performance, we measured the number of factual elements correctly included in the final answer, comparing runs with and without the self-correction engine enabled. The results speak for themselves: on these challenging numerical queries, the self-correction engine delivered about 60% relative increase in answer coverage over key facts.

Overall Answer Quality: We also measured the system’s impact using Criteria-Eval, a checklist-style benchmark that is purpose-built to evaluate Samaya’s financial Q&A capabilities. This benchmark consists of complex, reasoning-heavy questions spanning a broad range of financial use cases. With the self-correction engine enabled, Samaya Q&A system achieved a 10.5% relative improvement over the baseline, highlighting its value in demanding, real-world Q&A scenarios.

Why This Matters to Our Users

In high-stakes workflows, especially in domains like finance, where data completeness and factual accuracy are non-negotiable, users need to trust that their tools won’t just provide “an answer”, but the right one.

Samaya’s self-correction engine enhances that trust by:

  • Improving answer completeness: Ensuring complex and broad queries are answered in full, even when they involve subtle pieces of information; and delivering this without compromise on speed.
  • Boosting user confidence: Users can move faster and lean more heavily on Samaya for complex research tasks, knowing that the system is constantly monitoring and improving its answer quality.
  • Reducing manual overhead: With fewer gaps to catch manually, users spend less time validating results and more time making decisions.

More broadly, this innovation highlights Samaya’s core strength: a real-time, adaptive knowledge engine built for experts tackling complex, high-stakes data — delivering speed, coverage, and precision users can trust.

Contributors

Roberto Dessi, Yuhao Zhang, Mingyi Yang, Michele Bevilacqua and Ashwin Paranjape contributed to the design and implementation of the self-correction engine.

Roberto Dessi led evaluation efforts with help from Yuhao Zhang, Mingyi Yang and Michele Bevilacqua.

Special thanks to Christos Baziotis for his valuable discussions and feedback throughout the project.