AICI

Microsoft AICI allows to constrain the output of a LLM using WASM. In particular, it is possible to send JavaScript program to describe the prompt.

GenAIScript support executing scripts and converting the output into a AICI compatible JavaScript program, which will them generate constrainted output.

Let’s take a look at an example.

$`Ultimate answer is to the life, universe
and everything is ${AICI.gen({ regex: /\d\d/ })}`

The execution of this script is converted into a AICI JavaScript program.

async function aiciregex() {
    await $`Ultimate answer is to the life, universe and everything is `
    await gen({ regex: /\d\d/ })
}

async function main() {
    await aiciregex()
}
start(main)

And AICI comes back with the following log.

FIXED "Ultimate answer is to the life, universe and everything is "
GEN-OPT {regex: /\d\d/}
regex constraint: "\\d\\d"
dfa: 160 bytes

GEN "42"
JsCtrl: done

And the text output is 42.

Metadata

An AICI template should set the aici provider in the model identifier.

script({ ...
    model: "aici:mixtral",
})

`gen`

The AICI.gen function creates a constrain in the prompt flow.

Token

AICI uses AICI_API_KEY, AICI_API_BASE and AICI_API_VERSION (default v1) to compose the API URL.

<base>/<model>/<version>/run