Forty transcripts, one question.
The answer is somewhere in the prepared remarks and the Q&A of every one of them, phrased differently each time and occasionally not given at all. Ragextract reads the corpus once and returns the column, including the rows where the question was dodged.
Free credits on sign-up. No card, no subscription.
| Row | Documents | Pricing guidance | Launch dates | Headcount | Period |
|---|---|---|---|---|---|
| Kestrel Systems Q3.pdf | List prices held; discounting to normalise in Q4 p.6 | Platform v4 in Q1; hardware refresh mid-year p.9 | Hiring paused outside engineering p.11 | Q3 p.1 | |
| Harlow Group Q3.pdf | 3–4% increase from January p.4 | No launches mentioned | Restructuring, 200 roles p.8 | Q3 p.1 | |
| Pelham Industrial Q3.pdf | Declined to guide on pricing p.14 | New line commissioning in H2 p.7 | Headcount flat year on year p.5 | Q3 p.1 | |
| Kestrel Systems Q2.pdf | Discounting elevated, expected to persist p.5 | Platform v4 in Q4 p.8 | Hiring across all functions p.10 | Q2 p.1 |
A quote you can put in the note.
Research output lives or dies on attribution. A claim about what management said in Q3 is worth having only if the page it was said on travels with it, and a summary that cannot be traced back is a summary somebody else has to redo.
Every cell carries its page and a confidence score, so the checking goes to the answers that look thin rather than back through the corpus. What the table changes is the ratio of reading to writing, not whether the reading happens.
| Document | Sample size | Funding conflict |
|---|---|---|
| Cohort study, 2024 | 1,284 p.3 | Yes — industry-funded p.11 |
| Meta-analysis, 2025 | 17 studies, 9,402 participants p.2 | No p.14 |
A row is one subject, one period.
This is the decision that makes a corpus into a series. One company across four quarters is four rows, not one, and a period column is what lets the table be read down as well as across — which is where the interesting answer usually is.
Bundling is the exception rather than the rule here. A transcript and its accompanying slide deck are one event and belong together; two quarters of the same company are two events and do not. Bundle the first, never the second.
Set the output type to match what you will do with the answer. Sample size is a number so it can be sorted; launch dates are a list because there are usually several; “declined to guide” is a legitimate text answer and a much more useful one than a blank.
Two of the types answer with a picture rather than a sentence, which is what the literature case has always wanted. The Academia preset asks for every figure presenting a study’s results, with its axis labels and legend, and puts them in the row — so a methodological comparison across forty papers can be looked at rather than described.
Where it fits.
Two very different corpora, one job: find the handful of fields that make the members comparable, and put them side by side.
| The work | What a row is | Start from |
|---|---|---|
| Earnings and transcript work | One company, one quarter | |
| Literature and evidence review | A published paper | |
| Competitive and market scanning | A filing, report or release | |
| Longitudinal tracking | One subject, one period |
Both presets ship in the application, and both are the ones not yet worked through under use cases — so start from columns, which covers the output types and how to write a prompt that holds up, and take the closest worked setup for the shape.
Reading is charged once.
A corpus is billed for the pages it reads, not for the questions asked of them. Next quarter’s question, asked of last quarter’s transcripts, costs only its own column — which is the difference between a corpus you maintain and one you rebuild.
| One quarter, 40 companies | Volume | Cost |
|---|---|---|
| 40 transcripts, 30 pages each — read once | 1,200 pages | 1,200 credits |
| 3 typed questions asked of all 40 | 120 cells | 120 credits |
| Launch dates, as a list column | 40 cells | 120 credits |
Start with two quarters of one company.
Upload them, write the column you would have gone looking for by hand, and read both citations. Two rows of a company you already follow is the fastest way to find out whether the answers are the ones you would have written down yourself.
Nothing on this page is investment or research advice, and the presets are not methodological definitions.