|
| 1 | +# Vitest AI |
| 2 | + |
| 3 | +**Vai** (stands for _Vitest + AI_) is a lightweight vitest extension that uses LLMs to do assertions. |
| 4 | +The goal of this library is primarily to allow testing the output of LLMs like the new autonomous engine, as the output is dynamic and qualitative we can't rely on traditional hard-coded tests. |
| 5 | + |
| 6 | +To remove the flakiness and human-input from these tests, we need LLMs. |
| 7 | + |
| 8 | +It's built on top of Zui and the Botpress client to interface with the different LLMs. |
| 9 | + |
| 10 | +## Usage |
| 11 | + |
| 12 | +```typescript |
| 13 | +import { check, rate, filter, extract } from '@botpress/vai' |
| 14 | +import { describe, test } from 'vitest' |
| 15 | + |
| 16 | +describe('my test suite', () => { |
| 17 | + test('example', () => { |
| 18 | + check('botpress', 'is a chatbot company').toBe(true) |
| 19 | + }) |
| 20 | +}) |
| 21 | +``` |
| 22 | + |
| 23 | +## `check (assertion)` |
| 24 | + |
| 25 | +Checks that the provided value matches the provided condition |
| 26 | + |
| 27 | +```typescript |
| 28 | +test('example', () => { |
| 29 | + // works with strings |
| 30 | + check('hello', 'is a greeting message').toBe(true) |
| 31 | + |
| 32 | + // also works with objects, arrays etc.. |
| 33 | + check( |
| 34 | + { |
| 35 | + message: 'hello my friend', |
| 36 | + from: 'user' |
| 37 | + }, |
| 38 | + 'is a greeting message' |
| 39 | + ).toBe(true) |
| 40 | +}) |
| 41 | +``` |
| 42 | + |
| 43 | +## `extract (assertion)` |
| 44 | + |
| 45 | +Extracts from anything in input the requested Zui Schema: |
| 46 | + |
| 47 | +```typescript |
| 48 | +const person = z.object({ |
| 49 | + name: z.string(), |
| 50 | + age: z.number().optional(), |
| 51 | + country: z.string().optional() |
| 52 | +}) |
| 53 | + |
| 54 | +extract('My name is Sylvain, I am 33 yo and live in Canada', person).toMatchObject({ |
| 55 | + name: 'Sylvain', |
| 56 | + age: 33, |
| 57 | + country: 'Canada' |
| 58 | +}) |
| 59 | +``` |
| 60 | + |
| 61 | +Also added support for `toMatchInlineSnapshot`: |
| 62 | + |
| 63 | +```typescript |
| 64 | +test('toMatchInlineSnapshot', () => { |
| 65 | + extract('My name is Eric!', z.object({ name: z.string() })).toMatchInlineSnapshot(` |
| 66 | + { |
| 67 | + "name": "Eric", |
| 68 | + } |
| 69 | + `) |
| 70 | +}) |
| 71 | +``` |
| 72 | + |
| 73 | +## `filter (assertion)` |
| 74 | + |
| 75 | +Filters an array of anything `T[]` based on a provided condition: |
| 76 | + |
| 77 | +```typescript |
| 78 | +const countries = ['canada', 'germany', 'usa', 'paris', 'mexico'] |
| 79 | +filter(countries, 'is in north america').toBe(['canada', 'usa', 'mexico']) |
| 80 | +filter(countries, 'is the name of a country').length.toBe(4) |
| 81 | +``` |
| 82 | + |
| 83 | +## `rate (assertion)` |
| 84 | + |
| 85 | +Given any input `T`, gives a rating between `1` (worst) and `5` (best): |
| 86 | + |
| 87 | +```typescript |
| 88 | +test('good', () => rate('ghandi', 'is a good person').toBeGreaterThanOrEqual(4)) |
| 89 | +test('evil', () => rate('hitler', 'is a good person').toBe(3)) |
| 90 | +``` |
| 91 | + |
| 92 | +## Few-shot Examples |
| 93 | + |
| 94 | +All assertion methods accept examples to provide the LLM with few-shot learning and help increase the accuracy. |
| 95 | + |
| 96 | +```typescript |
| 97 | +describe('learns from examples', () => { |
| 98 | + test('examples are used to understand how to classify correctly', () => { |
| 99 | + const examples = [ |
| 100 | + { |
| 101 | + expected: true, |
| 102 | + value: 'Rasa the chatbot framework', |
| 103 | + reason: 'Rasa is a chatbot framework, so it competes with Botpress' |
| 104 | + }, |
| 105 | + { |
| 106 | + expected: false, |
| 107 | + value: 'Rasa the coffee company', |
| 108 | + reason: 'Since Rasa is a coffee company, it does not compete with Botpress which is not in the coffee business' |
| 109 | + } |
| 110 | + ] |
| 111 | + |
| 112 | + check('Voiceflow', 'is competitor', { examples }).toBe(true) |
| 113 | + check('Toyota', 'is competitor', { examples }).toBe(false) |
| 114 | + }) |
| 115 | +}) |
| 116 | +``` |
| 117 | + |
| 118 | +## Failure Messages |
| 119 | + |
| 120 | +All model predictions have nice failure messages by default: |
| 121 | + |
| 122 | +```typescript |
| 123 | +const countries = ['canada', 'germany', 'usa', 'paris', 'mexico'] |
| 124 | +filter(countries, 'is in north america').toBe(['canada', 'usa']) |
| 125 | +``` |
| 126 | + |
| 127 | +## Promises |
| 128 | + |
| 129 | +All assertion methods can also be used outside Vitest tests, as they return an `PromiseLike<T>` object that can be awaited. |
| 130 | + |
| 131 | +```typescript |
| 132 | +test('test truth', async () => { |
| 133 | + const { result } = await check('hello', 'is a greeting message') |
| 134 | + expect(result).toBe(true) |
| 135 | +}) |
| 136 | +``` |
| 137 | + |
| 138 | +## Bail on failure |
| 139 | + |
| 140 | +You can await the assertion to bail immediately on failure and prevent other test cases to run: |
| 141 | + |
| 142 | +```typescript |
| 143 | +test('no bail', () => { |
| 144 | + check('hello', 'is a greeting message').toBe(false) |
| 145 | + console.log('this will run as the above is not awaited, it will bail at the end of the test') |
| 146 | +}) |
| 147 | + |
| 148 | +test('bail', async () => { |
| 149 | + await check('hello', 'is a greeting message').toBe(false) |
| 150 | + console.log('this will not run, the test has bailed') |
| 151 | +}) |
| 152 | +``` |
| 153 | + |
| 154 | +## Changing the evaluator model |
| 155 | + |
| 156 | +By default, GPT-4o mini is used to evaluate the tests, but the evaluator can be changed from inside a test: |
| 157 | + |
| 158 | +```typescript |
| 159 | +test('simple', () => { |
| 160 | + setEvaluator('anthropic__claude-3-5-sonnet-20240620') |
| 161 | + rate('hello', 'is a greeting message').toBe(5) |
| 162 | +}) |
| 163 | +``` |
0 commit comments