Skip to main content

Textless Mode

When you create a corpus via the API or the Vectara Console UI, you have the option to not store the text, also known as a "textless" mode. This mode is useful when you have very sensitive text content. The text content becomes unrecoverable to Vectara or to any user who successfully queries and finds the document.

tip

Textless mode is optimal for use cases where the cost of any information leakage is very high. Vectara does encrypt documents.

What happens in textless mode?

Let's look at when it's appropriate to enable textless, what happens on the platform, and what benefits and limitations it brings.

When you enable textless on a corpus, Vectara discards the text content of the document immediately after it converts the text to a vector. At that point, the text is no longer recoverable. It also won't be returned in any Vectara APIs.

note

Vectara does retain any metadata that were supplied alongside the text, including document IDs. This retention lets you retrieve the document from a separate system of record based on the ID to show the metadata, and it also allows Vectara to perform any metadata-based filtering on the document.

Enable Textless mode

To enable textless mode, set the textless value to true under corpus:


curl -X POST \
-H "Authorization: Bearer abcefg..." \
-H "customer-id: 123456789" \
https://api.vectara.io:443/v1/create-corpus \
-d @- <<END;
{
"corpus": {
"id": 2,
"name": "Name of my corpus",
"description": "Description about my corpus",
"dtProvision": 0,
"enabled": false,
"swapQenc": false,
"swapIenc": false,
//enable textless mode
"textless": true,
"encrypted": true,
"encoderId": 1,
"metadataMaxBytes": 0,
"faissIndexType": "",
"customDimensions": [],
"filterAttributes": [
{
"name": "lang",
"description": "Detected language, as an ISO 639-3 code.",
"indexed": false,
"type": 25,
"level": 10
},
{
"name": "is_title",
"description": "True if the text is a title.",
"indexed": false,
"type": 35,
"level": 10
}
]
}
}
END

Limitations

Currently, the reranking capability relies on the text being stored. As a result, attempting to rerank search results on any corpora where text storage has been turned off will not work at this time.