Research workflow
What happens between pressing Enter and reading the answer. The steps below are the ones the run page shows. All five models follow the same outline and differ in a few steps.
01Starting a run
A run starts from the research box with four choices: the question, the model, where to look (Web, Papers or Auto), and the project to save it in. The Pro models also take a depth: Scan, Deep Dive or Exhaustive. See Models and plans.
Starting a run uses one of your monthly research runs. The run is saved to the project straight away, so it appears in the project and under Recent while it works.
02Understanding and planning
First the run reads the question. It works out what kind of question it is, its key concepts, the sub-questions worth answering, and whether it needs current information. It also classifies the topic, which decides how an Auto source search is split between papers and the web.
Planning turns this into search angles. On Auto, the model first picks a strategy (empirical, landscape or current) and says which one it chose.
On RA-1 and the models built on it, a step called Recalling strong sources may follow: sources that did well on the same topic in earlier runs are brought forward before the first search. It only appears when the topic has history.
03Searching
The search is steered by the model, round by round, the way a person searches:
- It writes a first round of queries, each aimed at a different angle.
- It looks at what came back: titles, sites and snippets.
- It chooses which results are worth opening, and reads them.
- With those results in view, it either decides it has enough or writes the next round.
The loop is bounded. On RA-Lite it runs at most four rounds and ten queries, with a fixed budget of pages to read. The other models have larger budgets set by the depth mode.
Web queries go to Brave Search, after a check of Sourcipedia's corpus of vetted sources. Papers queries go to six academic databases at once. Duplicate links are removed before anything is fetched. See Where sources come from.
Every query appears on the run page as it runs, with the results it returned, and short notes explain which pages the run chose to read and why.
04Reading pages
Each chosen page is fetched and reduced to its readable text. PDFs are read as well. When a site refuses automated readers, the run reads the Internet Archive's copy instead, and the source keeps its own address.
A page is set aside if it has fewer than 200 words of readable text, if it reads as navigation or garbled text, or if its body duplicates a page already kept. A few open-publishing sites are never used as sources. See Reading pages.
05Screening and choosing
Every readable source is scored. The signals are relevance to the question, the standing of the publisher, and recency. On RA-Lite, recency counts for more when the question is time-sensitive. RA-1 adds how much evidence a page carries, what kind of source it is, the domain's record across earlier runs, and a penalty for leaning too hard on one site.
The strongest sources are chosen for a deep read. A source that shares nothing with the question is never used to fill the set, so a thin search returns fewer sources rather than padding.
06Deep reading
Each chosen source is read closely: what it says about the question, its key findings, and quotes taken word for word. The reader also gives a verdict: the source answers the question directly, gives background, or has nothing on it.
A source with nothing on the question is taken out of the answer and listed under Other sources as read, with nothing on the question. The next-ranked source is read in its place. If every source is judged off-topic, the run keeps them all rather than answer from nothing.
07Writing and checking
The answer is written from the deep reads, with a marker such as [S1] after each claim pointing to its source. The model chooses the structure the question needs. An answer may also say where sources disagree and what remains uncertain.
- Citations. Each source gets a citation in six styles, built from its metadata. See Citations.
- Review. On RA-1 and the models built on it, Deep Dive and Exhaustive have a critic review the draft before it ships. Exhaustive also revises it.
- Evidence check. The answer's claims are checked against the evidence the run gathered. The share that held up is shown on the finished run.
- Citation check. RA-2 then checks every cited sentence against its source and fixes what fails.
- Gaps. Max then looks for what the sources missed, searches again, and revises the answer if new sources get in.
08The steps each model shows
The run page shows the current step as "Step N of M", with a short narration under it. There is no percentage bar. Steps that do not apply to a run are left out.
| Model | Steps |
|---|---|
| RA-Lite | Understanding the question · Planning the search · Searching · Removing duplicates · Reading pages · Screening sources · Choosing the strongest · Deep-reading sources · Preparing citations · Writing the answer |
| RA-1 | Understanding the question · Planning the search · Recalling strong sources · Searching · Pulling evidence · Screening sources · Deep-reading sources · Writing the answer |
| RA-2 | RA-1's steps, then Checking every citation |
| Max | First pass, RA-1's steps, then Filling the gaps and Deep-reading the rest |
| Auto | Opens with "Choosing how to research this…" and the strategy it picked, then runs RA-1's pipeline |
09The finished run
- Answer, with citation chips you can open.
- Sources the answer stands on, each with a trust label, the run's analysis and its citation.
- Other sources, found but not used. You can have up to three of them deep-read and added.
- Receipt: time taken, sources found, sources deep-read, and the share of claims backed by evidence. Open it to see the steps again.
- Copy and Export, for the answer as Markdown with its source list.
When a run finishes inside a project, the sources it stood on are saved to that project, so the project chat can read them. See Projects and documents.
10Stopping, failures and reopening
- Stop. Ends the run. The sources gathered so far stay on the page.
- Closing the tab. The run carries on on the server. Reopen it from Recent, its project, or Library → Runs, and the page catches up.
- Reopening later. A finished run is kept with its answer, references and every source, so it reads the same a month later.
- Writing fails. If the run gathered sources but could not write the answer, Try writing the answer again reuses those sources without searching again.
- Search fails. A run that finds no sources at all ends without an answer. Start it again.