Client reviews
“Send the review within three working days of receiving all materials, unless the project agreement specifies otherwise.”
← Research at Indriya
KNOWLEDGE & REASONING IN CONTEXT
Which answer can you use when your sources disagree?
Your team’s knowledge is spread across documents, conversations and experience. Finding a plausible answer is not the same as knowing which information applies.
We investigate how people can trace an explanation to its sources, recognize a gap, and decide what to use.
Explore a changing answer
A QUESTION IN CONTEXT
Change the evidence.
See why the answer changes.
A project lead asks: “When should we send the client’s review?” Change what is known and see why the answer needs to change with it.
WHAT THE SOURCE SAYS
All materials have arrived. No project-specific exception has been supplied.
SUPPORTED BY THE PLAYBOOK
Within three working days— unless this project has an exception.
The standard gives a starting point. The condition in the same sentence matters just as much as the deadline.
A useful answer preserves the qualification.
WHAT THE SOURCES SAY
Three working days
Unless the project agreement specifies otherwise.
Next working day
“For this project, send the review by the end of the next working day after all materials arrive.”
THE EXCEPTION APPLIES
The next working day. The project agreement changes the answer.
Its authority and scope matter—not simply its later date. The playbook explicitly allows this project-specific exception.
Retrieving both passages is not enough. Their relationship must remain visible.
WHAT IS KNOWN
“Timing changed on the call.”
The agreed deadline is not recorded in the material available to the team.
The standard playbook is available. The agreement that may override it is not.
CONFIRM BEFORE COMMITTING
The deadline is not established. Ask for the missing agreement.
Repeating the standard would conceal an unresolved exception. A helpful system should identify the gap and make it easy to resolve.
Sometimes the useful result is a better question.
WHY IT IS DIFFICULT
How do you decide
which source applies?
You may find more sources and still face conflicting answers. To decide what to use, you need to check what each source actually supports.
Which information should govern?
A general rule, a project exception and an informal note have different roles. We are interested in making scope, authority and unresolved conflicts explicit—not treating every retrieved passage as interchangeable.
Does the passage support the claim?
A reference may concern the right topic while missing the qualification that changes the answer. The research question is whether people can detect that mismatch without redoing the entire task.
When should the system ask?
Asking for every missing detail creates work. Guessing a consequential detail creates risk. The useful balance depends on the decision, its reversibility and what the person can verify.
WORK YOU CAN EXAMINE
Keep the explanation
close to its source.
Our Learning Studio prototype brings selected materials, passage references and recall activities into one workspace. It gives us a concrete setting in which to examine checking and understanding.
It is in development. The recorded example uses sample material and prewritten excerpts, not a live AI response. We have not established that it improves verification accuracy or learning outcomes.
How could we test this—and what work does it build on?
We would compare a bounded task against the participant’s current reading and checking workflow. Include supported claims, misleading references and missing evidence; examine corrections, task completion and effort together. A faster answer is not sufficient if important errors go unnoticed.
Task-based walkthroughs and source checks are described in our proposed evaluation plan. Participant recruitment, scoring and study design remain to be finalized; these are not completed findings.
Read the draft evaluation plan ↓Related reading: Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) describes combining retrieval with generation. It provides a foundation for examining how retrieved evidence is interpreted and checked in a particular task.
PUT YOUR QUESTION TO WORK
Where does your team
lose the context?
For a small business, it may be a handover that depends on one person’s memory. For an enterprise team, it may be conflicting guidance across approved sources. For a researcher, it may be how to measure whether an explanation is genuinely supported.
Bring a non-confidential example of the task and the information it depends on. We can explore the fit for a bounded prototype or evaluation, with data access and success criteria agreed before work begins.
Discuss a knowledge problemCONTINUE EXPLORINGAll research areas