AuraDoc
Sign inSign up

Blog

Semantic search across all your documents

Find documents by meaning, not just keywords: hybrid search, filters, search by example, saved searches and alerts across your whole library.


Keyword search finds the words you typed. That works until the document says termination and you searched for cancellation, or until a policy describes the same obligation in three different phrasings. Semantic search is built for that gap: instead of matching characters, it matches meaning, so a question in your own words can land on the passage that answers it.

How semantic search works in AuraDoc

When a document is processed, its text is split into passages and each passage is turned into an embedding, a numeric representation of what the passage is about. Those embeddings are stored in a vector index. When you search, your query is turned into an embedding in the same way and compared with the stored passages, and the closest ones come back as results. The API search endpoint returns ten results by default and up to fifty, ignores weak matches below a minimum score, and can re-rank the shortlist for a better ordering. Each result carries a snippet from the passage with the matching terms highlighted, so you can judge a hit without opening the file.

Search covers whatever text the extraction produced, so scanned files are included once they have been through the OCR pipeline. If a document has no text yet, there is nothing for the index to hold.

Keyword and meaning together

Meaning is not always what you want. Sometimes you need an exact invoice number or a clause reference, and a purely semantic match can drift. For this there is a hybrid search that combines keyword matching with the vector search. It is also the safety net: if no key is available for the embedding provider, the search does not fail with a provider error. It falls back to matching words, and the response includes a note that says why. Semantic search itself needs an embedding provider to be configured, either by the platform or through your own key in the settings. When none is available, the response says so instead of returning silent empty results.

Narrowing results with filters

A meaning-based search over a large library returns a lot, so filters matter. You can restrict a search to a single document, or limit it by file type such as PDF, Word, Excel, image or text, by upload date, by file name, by size range, by who uploaded the file and by processing status. You can also keep only documents that have extracted entities, signatures or detected personal data, and filter by where a file came from, for example a synced cloud folder. A separate field matches against the entities the AI found in a document, which is handy for questions such as which contracts mention a particular company. Results can be ordered by relevance, newest, oldest or file name.

Search by example

You do not need words at all if you already have a document. The search can start from a document and return the ones most similar to it, which is the same idea behind the similar documents tab on a document page. In the web app, the search page also accepts a link that starts from a given document, so you can share a starting point with a colleague. To go from a similar file to the exact differences, see Comparing two documents.

Saved searches and alerts

A question you ask every week should not be retyped. Save a search with a name, its query and its filters, and run it again with one action. A saved search can also notify you about new documents, and an alert can be set to fire immediately, daily or weekly. That turns search into a light monitoring tool: a saved search for a clause you never want to miss, checked as new files arrive from your watched cloud folders.

Cost and permissions

Each query is embedded by the provider, so the embedding counts against your quota, and the platform checks that the quota covers it before paying for it. API access uses a dedicated search read scope, so you can give an integration permission to search without giving it more. When you want answers instead of a list, try chat with citations. Plans and limits are on the pricing page.

Back to the blog