supabase/apps/docs/content/guides/functions/ai-models.mdx

---
id: 'function-ai-models'
title: 'Running AI Models'
description: 'How to run AI models in Edge Functions.'
subtitle: 'How to run AI models in Edge Functions.'
tocVideo: 'w4Rr_1whU-U'
---

[Supabase Edge Runtime](https://github.com/supabase/edge-runtime) has a built-in API for running AI models. You can use this API to generate embeddings, build conversational workflows, and do other AI related tasks in your Edge Functions.

## Setup

There are no external dependencies or packages to install to enable the API.

You can create a new inference session by doing:

```ts
const model = new Supabase.ai.Session('model-name')
```

<Admonition type="tip">

To get type hints and checks for the API you can import types from `functions-js` at the top of your file:

```ts
import 'jsr:@supabase/functions-js/edge-runtime.d.ts'
```

</Admonition>

## Running a model inference

Once the session is instantiated, you can call it with inputs to perform inferences. Depending on the model you run, you may need to provide different options (discussed below).

```ts
const output = await model.run(input, options)
```

## How to generate text embeddings

Now let's see how to write an Edge Function using the `Supabase.ai` API to generate text embeddings. Currently, `Supabase.ai` API only supports the [gte-small](https://huggingface.co/Supabase/gte-small) model.

<Admonition type="tip">

`gte-small` model exclusively caters to English texts, and any lengthy texts will be truncated to a maximum of 512 tokens. While you can provide inputs longer than 512 tokens, truncation may affect the accuracy.

</Admonition>

```ts
const model = new Supabase.ai.Session('gte-small')

Deno.serve(async (req: Request) => {
  const params = new URL(req.url).searchParams
  const input = params.get('input')
  const output = await model.run(input, { mean_pool: true, normalize: true })
  return new Response(JSON.stringify(output), {
    headers: {
      'Content-Type': 'application/json',
      Connection: 'keep-alive',
    },
  })
})
```

## Using Large Language Models (LLM)

Inference via larger models is supported via [Ollama](https://ollama.com/) and [Mozilla Llamafile](https://github.com/Mozilla-Ocho/llamafile). In the first iteration, you can use it with a self-managed Ollama or [Llamafile server](https://www.docker.com/blog/a-quick-guide-to-containerizing-llamafile-with-docker-for-ai-applications/). We are progressively rolling out support for the hosted solution. To sign up for early access, fill up [this form](https://forms.supabase.com/supabase.ai-llm-early-access).

<video width="99%" muted playsInline controls={true}>
  <source
    src="https://xguihxuzqibwxjnimxev.supabase.co/storage/v1/object/public/videos/docs/guides/edge-functions-inference-2.mp4"
    type="video/mp4"
  />
</video>

### Running locally

<Tabs
  scrollable
  size="large"
  type="underlined"
  defaultActiveId="ollama"
  queryGroup="platform"
>
  <TabPanel id="ollama" label="Ollama">

[Install Ollama](https://github.com/ollama/ollama?tab=readme-ov-file#ollama) and pull the Mistral model

```bash
ollama pull mistral
```

Run the Ollama server locally

```bash
ollama serve
```

Set a function secret called AI_INFERENCE_API_HOST to point to the Ollama server

```bash
echo "AI_INFERENCE_API_HOST=http://host.docker.internal:11434" >> supabase/functions/.env
```

Create a new function with the following code

```bash
supabase functions new ollama-test
```

```ts supabase/functions/ollama-test/index.ts
import 'jsr:@supabase/functions-js/edge-runtime.d.ts'
const session = new Supabase.ai.Session('mistral')

Deno.serve(async (req: Request) => {
  const params = new URL(req.url).searchParams
  const prompt = params.get('prompt') ?? ''

  // Get the output as a stream
  const output = await session.run(prompt, { stream: true })

  const headers = new Headers({
    'Content-Type': 'text/event-stream',
    Connection: 'keep-alive',
  })

  // Create a stream
  const stream = new ReadableStream({
    async start(controller) {
      const encoder = new TextEncoder()

      try {
        for await (const chunk of output) {
          controller.enqueue(encoder.encode(chunk.response ?? ''))
        }
      } catch (err) {
        console.error('Stream error:', err)
      } finally {
        controller.close()
      }
    },
  })

  // Return the stream to the user
  return new Response(stream, {
    headers,
  })
})
```

Serve the function

```bash
supabase functions serve --env-file supabase/functions/.env
```

Execute the function

```bash
curl --get "http://localhost:54321/functions/v1/ollama-test" \
--data-urlencode "prompt=write a short rap song about Supabase, the Postgres Developer platform, as sung by Nicki Minaj" \
-H "Authorization: $ANON_KEY"
```

</TabPanel>
<TabPanel id="llamafile" label="Mozilla Llamafile">

Follow the [Llamafile Quickstart](https://github.com/Mozilla-Ocho/llamafile?tab=readme-ov-file#quickstart) to download an run a Llamafile locally on your machine.

Since Llamafile provides an OpenAI API compatible server, you can either use it with `@supabase/functions-js` or with the official OpenAI Deno SDK.

<Tabs
  scrollable
  size="large"
  type="underlined"
  defaultActiveId="supabase-functions-js"
  queryGroup="sdk"
>
  <TabPanel id="supabase-functions-js" label="Supabase Functions JS">

Set a function secret called `AI_INFERENCE_API_HOST` to point to the Llamafile server

```bash
echo "AI_INFERENCE_API_HOST=http://host.docker.internal:8080" >> supabase/functions/.env
```

Create a new function with the following code

```bash
supabase functions new llamafile-test
```

<Admonition type="note">

Note that the model parameter doesn't have any effect here! The model depends on which Llamafile is currently running!

</Admonition>

```ts supabase/functions/llamafile-test/index.ts
import 'jsr:@supabase/functions-js/edge-runtime.d.ts'
const session = new Supabase.ai.Session('LLaMA_CPP')

Deno.serve(async (req: Request) => {
  const params = new URL(req.url).searchParams
  const prompt = params.get('prompt') ?? ''

  // Get the output as a stream
  const output = await session.run(
    {
      messages: [
        {
          role: 'system',
          content:
            'You are LLAMAfile, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests.',
        },
        {
          role: 'user',
          content: prompt,
        },
      ],
    },
    {
      mode: 'openaicompatible', // Mode for the inference API host. (default: 'ollama')
      stream: false,
    }
  )

  console.log('done')
  return Response.json(output)
})
```

</TabPanel>
<TabPanel id="openai" label="OpenAI Deno SDK">

Set the following function secrets to point the OpenAI SDK to the Llamafile server:

```bash
echo "OPENAI_BASE_URL=http://host.docker.internal:8080/v1" >> supabase/functions/.env
echo "OPENAI_BASE_URL=OPENAI_API_KEY=sk-XXXXXXXX" >> supabase/functions/.env
```

Create a new function with the following code

```bash
supabase functions new llamafile-test
```

<Admonition type="note">

Note that the model parameter doesn't have any effect here! The model depends on which Llamafile is currently running!

</Admonition>

```ts supabase/functions/llamafile-test/index.ts
import OpenAI from 'https://deno.land/x/openai@v4.53.2/mod.ts'

Deno.serve(async (req) => {
  const client = new OpenAI()
  const { prompt } = await req.json()
  const stream = true

  const chatCompletion = await client.chat.completions.create({
    model: 'LLaMA_CPP',
    stream,
    messages: [
      {
        role: 'system',
        content:
          'You are LLAMAfile, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests.',
      },
      {
        role: 'user',
        content: prompt,
      },
    ],
  })

  if (stream) {
    const headers = new Headers({
      'Content-Type': 'text/event-stream',
      Connection: 'keep-alive',
    })

    // Create a stream
    const stream = new ReadableStream({
      async start(controller) {
        const encoder = new TextEncoder()

        try {
          for await (const part of chatCompletion) {
            controller.enqueue(encoder.encode(part.choices[0]?.delta?.content || ''))
          }
        } catch (err) {
          console.error('Stream error:', err)
        } finally {
          controller.close()
        }
      },
    })

    // Return the stream to the user
    return new Response(stream, {
      headers,
    })
  }

  return Response.json(chatCompletion)
})
```

</TabPanel>
</Tabs>

Serve the function

```bash
supabase functions serve --env-file supabase/functions/.env
```

Execute the function

```bash
curl --get "http://localhost:54321/functions/v1/llamafile-test" \
 --data-urlencode "prompt=write a short rap song about Supabase, the Postgres Developer platform, as sung by Nicki Minaj" \
 -H "Authorization: $ANON_KEY"
```

</TabPanel>
</Tabs>

### Deploying to production

Once the function is working locally, it's time to deploy to production.

Deploy an Ollama or Llamafile server and set a function secret called `AI_INFERENCE_API_HOST` to point to the deployed server

```bash
supabase secrets set AI_INFERENCE_API_HOST=https://path-to-your-llm-server/
```

Deploy the Supabase function

```bash
supabase functions deploy
```

Execute the function

```bash
curl --get "https://project-ref.supabase.co/functions/v1/ollama-test" \
 --data-urlencode "prompt=write a short rap song about Supabase, the Postgres Developer platform, as sung by Nicki Minaj" \
 -H "Authorization: $ANON_KEY"
```

As demonstrated in the video above, running Ollama locally is typically slower than running it in on a server with dedicated GPUs. We are collaborating with the Ollama team to improve local performance.

In the future, a hosted LLM API, will be provided as part of the Supabase platform. Supabase will scale and manage the API and GPUs for you. To sign up for early access, fill up [this form](https://forms.supabase.com/supabase.ai-llm-early-access).