Sources and library
Where a run's sources come from, how they are read and judged, and where they are kept afterwards.
01Where sources come from
The Sources chip in the research box sets where a run looks. It can be changed on every plan.
Web
Web searches go to Brave Search. Before paying for a web search, the run checks the corpus of vetted sources behind Sourcipedia. When the corpus already covers what the search needs, the web search is skipped. When it covers part, the web search is smaller and is steered away from sites the corpus already supplied.
Papers
Papers searches six academic databases at once: Semantic Scholar, OpenAlex, Crossref, PubMed, arXiv and Europe PMC. Results are merged, duplicates are removed by DOI and address, and each paper is ranked by how well it matches the question and how reliable the database's records are. A paper has to share at least two meaningful words with the question to count as a match.
Auto
Auto, the default, searches both and weights the mix to the question. Scientific and academic questions give papers the larger share. Questions about companies, products, the news, law or policy lean almost entirely on the web, where the primary material is. Papers that barely match the question are held back and used only if the web cannot fill the pool.
02Reading pages
Duplicate links are removed before anything is fetched. Each page the run chooses is then fetched and reduced to its readable text.
- Web pages are cleaned of navigation and other page furniture.
- PDFs are read too, so statutes, court documents and agency reports can be used.
- Blocked pages. When a site refuses automated readers, the run reads the Internet Archive's copy if there is one. The source keeps its own address.
A page is not used if it has fewer than 200 words of readable text, if the text reads as navigation or garbage, or if its body duplicates a page already kept. A page that could not be read is shown as unavailable.
Pages from these open-publishing sites are never used as sources: medium.com, quora.com, blogspot.com, wordpress.com, tumblr.com and study.com.
03Scoring and selection
Every page that was read is scored, and the strongest are chosen for a deep read. The signals depend on the model.
- RA-Lite weighs relevance to the question, the authority of the site, how reputable it is, and recency.
- RA-1, and the models built on it, weigh relevance, credibility, recency, how much evidence the page carries and what kind of source it is. Credibility blends fixed signals about the site with its record across earlier runs. There is a small bonus for sources that did well on the topic before, and a penalty for relying on one site too often.
On RA-Lite, recency counts for more when the question is time-sensitive. A source that shares nothing with the question is never used to fill the deep-read set.
04Trust labels
Each source in a run carries a trust label. Click it for the details: what the source is authoritative (or useful) for, what it is not, and the signals behind the label.
| Kind of publisher | Label | Good for | Not for |
|---|---|---|---|
| Government agency | Trusted | Official statistics, regulations and public guidance | An outside review of the agency's own position |
| Intergovernmental body | Trusted | Cross-country statistics and policy guidance | The newest evidence, where consensus lags |
| Peer-reviewed publisher | Trusted | Peer-reviewed findings in its subjects | Settled consensus from a single study |
| Academic institution | Trusted | Its research and teaching material | Consistent rigour across every staff and course page |
| Preprint server | Established | What the authors claim, early | Peer-reviewed findings |
| Established newsroom | Established | What was reported and said | The underlying data; go to the primary source |
| General reference work | Established | Orientation, terms and finding the primary literature | Citing directly; cite what it references |
| An organisation's own site | Established | Its own announcements, documentation and figures | An independent check on its claims |
| Open-publishing platform | Use with care | Opinion and firsthand accounts | Anything unverified; there is no editorial review |
| Unknown site | Unverified | — | More than a single unverified account |
Once a site has enough history across runs, how often its claims held up can move its label up or down. History never changes what a kind of source is good for. Labels are worked out from the address and that history. No AI model is involved.
05Deep reads and other sources
Each chosen source is read closely. Open it in the run's source list to see what it says about the question, its key findings and quotes.
The reader can also decide a source has nothing on the question. That source is taken out of the answer and the next-ranked source is read in its place. It still appears under Other sources, marked as read with nothing on the question, so you can tell it apart from sources the run never opened.
Other sources is collapsed by default. It lists everything the run found but did not reason from. If you spot one worth reading, you can have up to three of them deep-read and added to the set, without starting a new run.
06Saving sources
When a run in a project finishes, the sources the answer stands on are saved to that project with their analysis, text and citation. The project chat can then read them and cite them by paragraph.
You can add more to a project yourself:
- paste a link into the Sources drawer in the workspace,
- paste a link into the project chat's message box; it is saved when you send,
- type
/addsourceand a link in the chat, - ask the chat to add a source.
An added link is read and analysed before it is saved. A project's Sources page lists everything saved to it. You can search by title or site, sort, copy a link, or remove sources.
07Your library
Saved sources belong to your account's library, and projects point to them. The same source can sit in several projects. Removing a source from a project takes it out of that project only; it stays in your library.
Each plan has a cap on how many distinct sources the library holds: 50 on Free, 500 on Pro and 5,000 on Researcher. Saving a link you already have does not use another slot. At the cap, new sources are not saved until you move to a larger plan.
Library in the sidebar lists your projects, with how many sources and chats each has, and every run you have started.