The goal of this project is to create a **set of standard AI functions / tools** which are **optimized for both normal TS-usage as well as LLM-based usage** across any popular AI SDK via simple adaptors.
For example, all of the stdlib tools like `WeatherClient` can be used both as normal, fully-typed TS SDKs:
```ts
import { WeatherClient } from '@agentic/stdlib'
const weather = new WeatherClient() // (requires `WEATHER_API_KEY` env var)
const result = await clearbit.getCurrentWeather({
q: 'San Francisco'
})
console.log(result)
```
Or you can use them as a set of LLM-based functions / tools using adaptors specific to each LLM SDK. This example uses [Vercel's AI SDK](https://github.com/vercel/ai):
```ts
// sdk-specific imports
import { openai } from '@ai-sdk/openai'
import { generateText } from 'ai'
import { createAISDKTools } from '@agentic/stdlib/ai-sdk'
// sdk-agnostic imports
import { WeatherClient } from '@agentic/stdlib'
const weather = new WeatherClient()
const result = await generateText({
model: openai('gpt-4o'),
tools: createAISDKTools(weather),
toolChoice: 'required',
prompt: 'What is the weather in San Francisco?'
})
console.log(result.toolResults[0])
```
Let's take a slightly more complicated example which uses multiple clients and selects a subset of their functions using the `pick` method:
```ts
// sdk-specific imports
import { ChatModel, createAIRunner } from '@dexaai/dexter'
import { createDexterFunctions } from '@agentic/stdlib/dexter'
// sdk-agnostic imports
import { PerigonClient, SerperClient } from '@agentic/stdlib'
async function main() {
const perigon = new PerigonClient()
const serper = new SerperClient()
const runner = createAIRunner({
chatModel: new ChatModel({
params: { model: 'gpt-4o', temperature: 0 }
// debug: true
}),
functions: createDexterFunctions(
perigon.functions.pick('search_news_stories'),
serper
),
systemMessage:
'You are a helpful assistant. Be as concise as possible. Respond in markdown. Always cite your sources.'
})
const result = await runner(
'Summarize the latest news stories about the upcoming US election.'
)
console.log(result)
}
```
Here we've exposed 2 functions to the LLM, `search_news_stories` which corresponds to the `PerigonClient.searchStories` method and `serper_google_search` via the `SerperClient.search` method.
All of the SDK adaptors like `createDexterFunctions` are very flexible in what they accept, which is why we can pass an `AIFunctionSet` (like `perigon.functions.pick('search_news_stories')` or `perigon.functions` or `serper.functions`) or an `AIFunctioProvider` (like `perigon` or `serper`) or an individual `AIFunction` (like `perigon.functions.get('search_news_stories')` or `serper.functions.get('serper_google_search')`). You can pass as many or as few of these `AIFunctionLike` objects as you'd like and you can manipulate them as `AIFunctionSet` sets via `.pick`, `.omit`, and `.map`.
The SDK-specific imports are all isolated to keep the main `@agentic/stdlib` as lightweight as possible.
## Goals
- clients should be as minimal as possible
- clients must use `ky` as a lightweight native fetch wrapper
- clients must have a strongly-typed TS DX
- clients should expose select methods via the `@aiFunction(...)` decorator
-`@aiFunction` methods must use `zod` for input schema validation
- it should be easy to create external clients which follow the `AIFunctioProvider` superclass / `@aiFunction` decorator pattern
- common utility functions for LLM-based function calling should be exported for convenience
- clients and AIFunctions should be composable via `AIFunctionSet`
- clients must work with all major TS AI SDKs
- SDK adaptors should be as lightweight as possible and be optional peer dependencies of `@agentic/core`
- SDK adatptor entrypoints should all be isolated to their own top-level imports
