llmem 20240215

Read through cache for OpenAI's API.

llmem

A read through cache for OpenAI chat completions.

Why:

• Avoid running up bills when developing/testing by sending the same request multiple times/forgetting to mock things.
• Fine-tune an open source model on data I've paid for Adhere to all terms of use.
• Evaluate future models.
• Track model changes over time.

use

It currently only supports the chat completion API.

requests

NOTE: These examples run against my public server llmem.com. You might want to run your own server instead.

import requests

headers = {
  "Content-Type": "application/json",
  "Authorization": f"Bearer {api_key}"
}

payload = {
  "model": "gpt-3.5-turbo",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Hi."
        },
      ]
    }
  ],
  "max_tokens": 10
}

response = requests.post("https://llmem.com/chat/completions", headers=headers, json=payload)

print(response.json())

If you run that twice, you'll see that first it gives an ID starting with chatcmpl, then IDs starting with llmem.

You can also use the OpenAI client, although it will currently only work for /chat requests:

from openai import OpenAI

client = OpenAI(
    api_key=api_key,
    base_url="https://llmem.com"
)

chat_completion = client.chat.completions.create(
    messages=[
        {
            "role": "user",
            "content": "Say this is a test",
        }
    ],
    model="gpt-3.5-turbo",
)

private server

docker

Makes it available on port 8000, and keeps its cache in ~/.llmem.

docker run --rm -p 8000:8000  -v ~/.llmem:/workspace ghcr.io/c0g/llmem:latest

python

pip install .
uvicorn llmem:app

By default it will make a llmem.db file in the working directory, you can override this location by setting the LLMEM_DB_FILE environment variable.

You can get the code on GitHub.

public server

I host a public server at llmem.com. If you point your API client to https://llmem.com/chat, you'll hit my cache and share the wealth (and get to share other people's wealth).

• NOTE: I could look at your OpenAI key if I wanted to. I don't, but I could. You can probably trust me[^criminal].
• NOTE: I can look at your queries. I probably will (see 'Why' bullets above). Please don't submit anything you wouldn't want your mother seeing, since I might well be your mother.
• NOTE: If any of the above NOTEs worry you it's probably best not to point your API client to https://llmem.com.

[^criminal]: That's exactly what a criminal would say.