Skip to content

Use cases

AI for research

By Ofrai Editorial (Use Case Research)

Research is reading with a citation. Ofrai puts Claude Sonnet 5, GPT-5.3 Codex and Gemini 3.1 Pro on the same balance, so the reading pass and the synthesis pass can run on the same page of sources.

Open studio

Chat and video use credits from your Ofrai balance.

Reading a document

Ask for a claim-plus-location extract rather than a summary, and the difference is immediate. A summary tells you what the document seemed to be about; a claim list tells you what it asserted and where, which is the only part of reading you can actually check. Put the source open next to the output the entire time, because the fastest way to lose trust in a research assistant is to discover one confident sentence that the document never said.

Evidence with page references

Require every extracted claim to carry a page or section reference, and require the model to say it cannot find support when it cannot. Models will confidently produce a claim with no locator, and a claim without a locator cannot survive a review. The instruction to state when the document does not support something is the single most valuable sentence you can add, because it converts a fluent answer into a checkable one.

Comparing sources

Put the sources in one question rather than summarising them one at a time and hoping the differences emerge. Ask where two sources disagree, what each assumes that the other does not, and what would have to be true for both to hold. Agreement between sources is weaker evidence than it looks, because they often share the same original sources, and asking for shared assumptions surfaces that faster than any summary.

When to verify yourself

Verify anything you intend to build on. The model is a reading assistant, and reading assistants are good at finding plausible statements in a document and less good at noticing that the document was itself wrong, partial, or written to sell something. Treat the output as a map of where to look, then go and look. Anything that will end up in a decision, a recommendation, or a citation needs a human eye on the original source, not on the summary of it.

AI for research: a week of one literature review

Treat the first week as building a pipeline rather than finishing a review. On day one, run one document you have already read through the model and compare its extract against your own notes; the differences tell you which kind of document the model reads badly. On day two, write the extraction template for the kind of document that dominates your field, including the instruction to admit when a claim is unsupported, and test it on a second document.

On day three, take three documents that disagree with each other and ask for the disagreements, because that single question produces more usable structure than three separate summaries do. On day four, add the source-agreement check, asking what each source assumes, and note where two papers trace back to the same underlying study; this is where most spurious consensus in a literature comes from.

On day five, draft the synthesis section yourself with the extracts beside you rather than asking for a synthesis directly, because the model tends to flatten disagreement into false agreement when it is asked to write the summary. Over the week you end up with a set of templates, a list of the document types that need manual handling, and a working order of operations. The templates are the thing that pays off on the next review, because the second one should not cost a week.

Prompt patterns for evidence work

Four patterns carry most of the value. The extract pattern: 'List the claims this document makes about {topic}, each with the page or section it appears in and a short quotation, and say so explicitly if a claim cannot be located.' The disagreement pattern: 'These sources disagree on {question}. State what each one claims, what each one assumes, and which of them has direct evidence for it.' The assumption pattern: 'What would have to be true for this conclusion to hold? List the assumptions, not the arguments.'

The absence pattern, which is the one people skip and the one that catches fabrication: 'List the questions this document raises but does not answer, and where a reader should look for them.' The absence pattern is especially useful because a research model will rarely volunteer a gap, but it will enumerate one when asked directly, and that list is often the most valuable part of the reading.

Keep the variable slots visible so each pattern can be reused across a folder, and keep the source open next to the output while you work through it.

Choosing between Sonnet, Codex and Gemini for research

Sonnet is the reading engine and should be your default. It handles the ordinary case of a document making a set of claims that you need located, and it does it without making the answer hard to follow. Codex is the structure engine: when the output has to be a table, a JSON record or a comparable row across thirty documents, it is the one to use, and research projects usually need that more often than people expect once a folder grows.

Gemini 3.1 Pro is the long-context pass, and it earns its place when the question is genuinely cross-document, because you can give it several sources at once and ask for the synthesis rather than assembling it from separate extracts. The sensible division is to read with Sonnet, structure with Codex, and synthesise with Gemini when there are enough sources to justify holding them together.

The mistake to avoid is the opposite one, sending a single document to the long-context model because it sounds more capable and getting an answer you cannot trace back to a specific place.

Where AI research goes wrong

Four failures recur. The first is misattribution: the finding is real, but it is attached to the wrong paragraph, which means you will chase a source that does not say what you think it says. The second is confident invention: the document does not address the question and the model answers anyway, and the invented answer is usually the kind of thing the document would have said if it had.

The third is table drift, where a number that was correct in the prose comes out wrong in the tabular extract, and everyone downstream trusts the table because tables look checked. The fourth is the silent loss of a qualifier: 'in most trials' or 'for enterprise accounts' disappears, and the surviving claim is stronger than the source, which is how a hedged finding turns into a confident one.

The defences follow directly: always require a locator, always ask explicitly what the document does not address, check the numbers in the table against the prose, and read the sentence around every qualification rather than the conclusion it sits in.

AI for research: credits per source, not per project

Research tools tend to price by project or by seat, so reading the second paper feels like it should be free and often is not. On Ofrai the charge follows the reading itself, which means the economics of a folder are simple: each document costs what it costs, and re-reading is cheap relative to the first pass. Two consequences matter in practice.

First, the marginal cost of checking your own extract against the source is small, so you should always do it rather than trusting the model on the points that matter. Second, because the same balance covers the writing and the rest of the studio, a research week that ends in a draft does not need a second budget for the drafting. The model you use still matters, but the amount you spend is under your control in a way it is not with a per-seat subscription.

What to take into the draft, and what to verify

The output splits into a file of located claims and a file of open questions, and the two are worth different amounts of trust. The located claims with quotations are the material you can write from, provided you have opened the source at that location yourself at least once. The open questions are the most valuable output of the whole session, because a well-framed question about what a literature has not addressed is often a better contribution than another summary of what it did.

What you must verify yourself is everything that will be load-bearing: the numbers, the attributions, and the claims that will be quoted back to you. Keep the source open next to the draft while you write, and treat any sentence you cannot locate in the source as a sentence you do not yet have.

Who should build this in, and who should not

Build this in if your reading volume is large relative to your reading speed, which is common in policy, compliance, competitive intelligence, technical due diligence and any field where a new paper lands on you faster than you can read it. You will get the most from it in the parts of research that are mechanical: locating claims, building comparison tables, finding what a document did not say.

Stay out of it if your contribution is the reading itself, if the sources are short enough that reading them is faster than prompting, or if your field requires judging source credibility in ways the model will confidently smooth over. A cautionary note on the last one: the model cannot tell a peer-reviewed result from a marketing blog post, and it will treat them identically unless you tell it not to. The tool accelerates reading.

It does not replace judgement about what is worth reading, or whether it should be believed.

A first session with AI for research

Start with a document you have already read closely, so that you have an answer key. Ask for the claim-and-location extract with the instruction that unsupported claims must be flagged, and work through the output against the original.

You are looking for the same two things you would look for in any research tool: the places it found a detail you had missed, which is the case for keeping it, and the places it produced a confident claim that the document does not support, which is the case for tightening the instruction. Once the extract holds up on a document whose contents you know, take two documents that disagree and ask for the disagreements rather than for summaries.

That second test matters more than the first, because it is the closest thing the model will do to your actual work. If the disagreements come back clean, you have a research assistant; if they come back flattened, you have a summariser, and you should know which one you are holding before you build a workflow on it.

Open studio

FAQ

Can it cite the source?
It can give a page or section reference with a short quotation, which is the form that is actually checkable. What it will not do is guarantee the reference is right, so treat every locator as a pointer to verify rather than a citation to paste. A page reference you did not open is not a citation, it is a hope.
What happens if the document does not answer the question?
By default, you get an answer anyway, and it will be the kind of answer the document would have given if it had covered the topic. That is why the instruction to state when a claim cannot be located matters so much: it turns the model from a confident source into an honest one, and an honest not-found is more useful to you than a plausible invention.
How much does reading a document cost?
It depends on the document length and which model you point it at, and the live pricing is shown on the model pages in the studio. The practical approach is to read with the everyday model, structure with the one built for table output, and reserve the long-context model for the point where you genuinely need several sources held together.
How do I start on Ofrai?
Open a document you have already read, use the chat-with-PDF tool, and ask for claims with page references and an explicit instruction to flag anything it cannot locate. Work the output against the original. That is one session, and it tells you whether the tool is worth building a folder workflow on.
What is AI for research on Ofrai?
One workspace with Claude Sonnet 5, GPT-5.3 Codex, Gemini 3.1 Pro and a chat-with-PDF tool, on a single credit balance. Reading, structuring and synthesising all run in the same place, and you pick the model per document rather than per subscription.

Related models

Tools