How search works

Both modes start the same way and differ in how widely they search and how much they read. This is the same search the Sourcely app runs.

1. Your text becomes search phrasesCopy Anchor Link

A language model reads your whole text and writes Google Scholar search phrases from it, most relevant first:

• Texts under 22 words get three paraphrases of the text itself.
• Longer texts get three topical keyword phrases of three to nine words each.

The search uses the first two phrases, or the first three when the text is longer than 3000 characters. They are returned in document.search_queries, so you can show users what was searched. The phrases are written in the language of your text.

2. Quick searchCopy Anchor Link

For each phrase, quick search fetches one Google Scholar results page of up to 10 sources. If Google Scholar returns nothing, Semantic Scholar is used instead.

Results keep the provider's own ranking. Phrase one's results come first, then phrase two's, so a typical page holds up to 20 sources, or 30 for long texts. Duplicates and sources that fail your filters are removed, so a page can be shorter.

With the default medium enrichment, each returned source gets a relevance rating, a short analysis and quoted evidence, based on its abstract.

3. Deep searchCopy Anchor Link

Deep search runs five stages and usually takes about three minutes:

1. Search widely. For each phrase it collects up to 15 results from Semantic Scholar and three Google Scholar pages, up to 30 more results.
2. Shortlist. A language model reads every candidate's title and abstract against your text and picks the 15 most promising.
3. Read. Each shortlisted source's landing page and PDF are fetched and converted to text, three at a time.
4. Analyze. Each source is rated and analyzed against your text using its full text. The extracted text is returned in parsed_source.
5. Rank again. A second ranking pass orders the 15 sources using the analysis: relevance rating first, then study type, venue, citations and recency.

The default enrichment for deep search is high. With none or low, the shortlist order is returned without the final ranking.

Ordering and scoresCopy Anchor Link

document.source_ids is the final order. document.scores gives each source a rank score: larger is better, and the scores are not confidence percentages. Sources from preferred domains move ahead of the rest without reshuffling either group.

ClaimsCopy Anchor Link

With granularity: ["claim"], the API also picks citation-worthy sentences from your text and searches each one separately, up to max_claims. Each searched claim is one search phrase, so it returns one provider page in quick mode or one 15-source shortlist in deep mode.

CitationsCopy Anchor Link

Citations in APA 7, MLA 9, IEEE, Harvard and BibTeX are formatted from the source's metadata. They cost nothing extra. Check them against the original source before publishing.

What each mode costsCopy Anchor Link

You pay for the sources actually returned. See Credits and pricing for the full price list and subscription discounts.