GCP is Bullshit and Here's Why
Mon Jul 15, 2024You may have heard of these things called LLMs. And you might wonder how to go about using them in your various projects. This is an excellent thing to wonder about! I thoroughly recommend exploring it!
Calling LLMs
There's a company called OpenAI that has some endpoints for you to check out! Here's how you do it in Python:
import requests
resp = requests.post(
f"https://api.openai.com/v1/chat/completions",
headers={
"Content-Type": "application/json",
"Authorization": f"Bearer {your OpenAI API key}",
},
json={"messages": [{"role": "user", "content": "Hello there, OpenAI"}], "model": "gpt-4o"},
)
print(resp.json()["choices"][0]["message"]["content"])
Some people have begun pointing out that OpenAI is less than scrupulous in various ways, so we should be using other LLM API providers. That's a perfectly fine perspective! Anthropic is a competitor that has, as far as I know, a higher standard for AI safety and employee treatment. Here's how you use their system:
import requests
resp = requests.post(
f"https://api.openai.com/v1/messages",
headers={
"Content-Type": "application/json",
"x-api-key": ENV["ANTHROPIC_API_KEY"],
"anthropic-version": "2023-06-01",
},
json={"messages": [{"role": "user", "content": "Hello there, Anthropic"}], "model": "claude-3-5-sonnet-20240620"},
)
print(resp.json()["content"][0]["text"])
Ok, the response path is a bit more byzantine in the OpenAI case, and if you want to use system prompts in the Anthropic case, it's a bit weirder since it's a top-level parameter instead of a message type, and also Anthropic uses a bizarre x-api-key header instead of the web-standard Authorization, but you know what both of these have in common?
- You get an API key from your account
- You use that API key to authenticate your results
- You get a response back from the given model if you properly authenticate with your key
This Is The Part Where GCP Is Bullshit
Take a guess about how you do the comparable thing over on GCP.
Did you assume (after you set up a GCP project, then separately enable the Vertex API and generate an API key of course, can't take care of any of the bookkeeping automatically for you, what do you think this is, the glorious robot future?) that you could just, I dunno
import requests
requests.post(
f"https://us-central1-aiplatform.googleapis.com/v1/projects/{YOUR_PROJECT_ID}/locations/us-central1/publishers/google/models/gemini-1.5-flash-001a:streamGenerateContent",
headers={
"Content-Type": "application/json",
"Authorization": f"Bearer: {API_KEY}",
},
json={
"contents": {
"role": "user",
"parts": [
{
"text": "Hi there, GCP!"
}
],
}
},
)
Hah! You poor, sweet, summer child. That would never justify the kind of giant brains that Google API engineers like to pretend they have! What actually happens if you do that is:
{'error': {'code': 401,
'details': [{'@type': 'type.googleapis.com/google.rpc.ErrorInfo',
'domain': 'googleapis.com',
'metadata': {'method': 'google.cloud.aiplatform.v1.PredictionService.Predict',
'service': 'aiplatform.googleapis.com'},
'reason': 'CREDENTIALS_MISSING'}],
'message': 'Request is missing required authentication credential. '
'Expected OAuth 2 access token, login cookie or other '
'valid authentication credential. See '
'https://developers.google.com/identity/sign-in/web/devconsole-project.',
'status': 'UNAUTHENTICATED'}}
Wait, unauthenticated? But didn't we just use the API key we established?
Yes, but you see, you can't use API keys to access the Vertex API. For reasons. Meaning: that that API key you set up? Scoped only to access the Vertex API? It's completely useless. They let you do that for shits and giggles, because they hate you and your suffering is nectar to them.
Ok, how do we do this then? Oh, it's really simple! You just use the vertexai library! I mean, it's not as straightforward as calling requests.post, and it has a minor amount of vendor lock-in as a result, but we can abstract that later.
import vertexai
from vertexai.generative_models import GenerativeModel
vertexai.init(project="<your project id here>", location="us-central1")
model = GenerativeModel(model_name="gemini-1.5-flash-001")
response = model.generate_content(
"What's a good name for a flower shop that specializes in selling bouquets of dried flowers?"
)
That gets us up and running, right?
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/project/lib/python3.12/site-packages/vertexai/generative_models/_generative_models.py", line 407, in generate_content
return self._generate_content(
^^^^^^^^^^^^^^^^^^^^^^^
File "/project/lib/python3.12/site-packages/vertexai/generative_models/_generative_models.py", line 496, in _generate_content
gapic_response = self._prediction_client.generate_content(request=request)
^^^^^^^^^^^^^^^^^^^^^^^
File "/project/lib/python3.12/site-packages/vertexai/generative_models/_generative_models.py", line 210, in _prediction_client
aiplatform_initializer.global_config.create_client(
File "/project/lib/python3.12/site-packages/google/cloud/aiplatform/initializer.py", line 499, in create_client
"credentials": credentials or self.credentials,
^^^^^^^^^^^^^^^^
File "/project/lib/python3.12/site-packages/google/cloud/aiplatform/initializer.py", line 341, in credentials
self._set_project_as_env_var_or_google_auth_default()
File "/project/lib/python3.12/site-packages/google/cloud/aiplatform/initializer.py", line 98, in _set_project_as_env_var_or_google_auth_default
credentials, _ = google.auth.default()
^^^^^^^^^^^^^^^^^^^^^
File "/project/lib/python3.12/site-packages/google/auth/_default.py", line 691, in default
raise exceptions.DefaultCredentialsError(_CLOUD_SDK_MISSING_CREDENTIALS)
google.auth.exceptions.DefaultCredentialsError: Your default credentials were not found. To set up Application Default Credentials, see https://cloud.google.com/docs/authentication/external/set-up-adc for more information.
Fuck you.
Every single piece of documentation I've seen on this is highly insistent that you have to call gcloud auth application-default login in order to set up your credentials. The only problem with this is that it forces you to go through an OAuth authentication process with your Google account. As in, in-browser. Which means there is no obvious way to translate that into a deployable web-app.
I'll save you some searching, tweaking, testing and apoplectic rage: What you actually have to do is go back to that API authentication page and set up something called a Service Account for your project, granting it scopes to VertexAI. Then you have to download a JSON object corresponding to it that looks something like
{
"type": "service_account",
"project_id": "your project ID",
"private_key_id": "lol-nope",
"private_key": "-----BEGIN PRIVATE KEY-----\nlol-even-more-nope\n-----END PRIVATE KEY-----\n",
"client_email": "something@developer.gserviceaccount.com",
"client_id": "0000000000000000000000000000",
"auth_uri": "https://accounts.google.com/o/oauth2/auth",
"token_uri": "https://oauth2.googleapis.com/token",
"auth_provider_x509_cert_url": "https://www.googleapis.com/oauth2/v1/certs",
"client_x509_cert_url": "https://www.googleapis.com/robot/v1/metadata/x509/0000000000000000000000000000-compute%40developer.gserviceaccount.com",
"universe_domain": "googleapis.com"
}
Then you have to call
CREDS, PROJECT_ID = google.auth.load_credentials_from_file("wherever-you-saved-it.json", scopes=["https://www.googleapis.com/auth/cloud-platform"])
then you have to call
vertexai.init(project="<your project id here>", location="us-central1", credentials=CREDS)
to set up the Vertex credential provider background state. And if this is starting to sound like one of those Java Enterprise Class Name jokes to you too, congratulations, you might be as old as I am.
Now that all that's done, you can finally
model = GenerativeModel(model_name="gemini-1.5-flash-001")
response = model.generate_content(
"What's a good name for a flower shop that specializes in selling bouquets of dried flowers?"
)
print(response.text)
and get something approaching what you wanted
Here are some names for a flower shop specializing in dried flowers, playing on different themes:
**Nature & Botanical:**
* Whispering Wildflowers
* Sun-Kissed Blooms
* Everlasting Bloom
* The Dried Garden
* Bloom & Preserve
* The Botanical Bloom
* Dried & Divine
* Petals & Past
* Meadow & Muse
**Rustic & Vintage:**
* The Dried Bouquet
* Gather & Bloom
* The Vintage Petal
* The Bloom & Co.
* Dried & Charming
* The Willow & Rose
* The Timeless Bloom
* Dried & Loved
* The Floral Archive
**Unique & Playful:**
* Petal & Paper
* The Dried Petal Co.
* Forever & Bloom
* The Bloom Box
* The Dusty Petal
* The Dried Flower Studio
* Bloom in Time
* The Petal Post
**Tips for Choosing a Name:**
* **Keep it short and memorable:** A shorter name is easier for customers to remember and say.
* **Consider your target audience:** What kind of vibe do you want to convey? Rustic? Elegant? Whimsical?
* **Check for availability:** Make sure the name isn't already taken and that the domain name is available.
I hope this list helps you find the perfect name for your dried flower shop!
Is part of this me not bothering to wade through overtly complicated, sometimes contradictory, documentation and read deeply? Sure.
But remember those two code snippets from the beginning of this piece that showed you how to instantly and easily get results out of OpenAI and Anthropic? Please explain why it has to be more complicated than that. Please tell me why, in a world where literally all you're trying to do is get textual output from a particular model, you need to worry about the security model that Google (or more realistically AWS, which was then copied by Google) has arbitrarily decided to put in your way? If the Vertex API doesn't accept API keys, why is it even possible for you to create an API key scoped down to only the Vertex API without so much as a warning?
The complexity doesn't end at the above call in development by the way.
Suppose you wanted to put together a web service that makes this call when requested. How would you go deploying one that called into Anthropic? You'd build put together a server, and possibly Dockerfile, that checks the env var ANTHROPIC_API_KEY to figure out what to put in headers, and you're basically done. Every deployment secret manager in the world can handle this workflow easily and transparently, and if you don't like any of them, you could easily locally encrypt an .env file and give your build system the corresponding key.
What do you have to do with google? Oh, yeah. Here's a JSON file you need to load up. Granted, there's a google.auth.load_credentials_from_dict so you don't need to write it to disk when setting up creds, but
- that secret now has to contain a block of JSON,
- which you then need to load,
- and use to create a set of credentials,
- which you then need to use to initialize a credential manager,
- which you then need to have running in memory when you call
GenerativeModel(or possibly,model.generate_content, I'm still not entirely clear on it) - so that you can then follow the one single step you were actually here for and that every other AI API provider lets you skip to
A Reasoned Summary
Why is this bad?
I can foresee the counters that look like "well, if you just use their library, it becomes a non-issue", or "well, if you just call gcloud, there's no problem". The complaints I have about deployment complexity above probably have things of a similar shape that you could say.
Do you remember that joke about the lizards? Sorry, of course not, that clip is older than your dad. Go ahead and watch it, and then consider the analogy.
The thing is, sending an HTTPS POST request to an endpoint, accompanied by an authenticating API key is secure. It's about as secure as you can get. If you've got an attacker that's compromised your line to the point where they can exploit that opening, then you effectively auto-rotating keys every hour by deferring to the Google auth system buys you almost nothing. This is "more secure" in the sense that you might now get away from an attacker that has managed to compromise your line for less than an hour. Granted that's a gain, but it's a gain of somewhere between 0.001% and 1% additional security on top of the existing machinery. I highly doubt you could argue it up to more than that. This is a piddling amount of security, given that the other end of the trade is the balooning of development and deployment complexity I've outlined above.
This is Not. Worth. It. Point blank. If you made that trade, then you chose very poorly indeed. And I will rightly point at you and laugh if you try to defend it with any argument other than "sorry; I had like half an hour to put together the entire Vertex API schema and I was sick that day, and it was also the first API I designed, and also a dog was chewing on my gender-non-specific balls at the time so I was pretty distracted".
I also notice that you're banking on enough people making this trade and effectively getting locked into your platform down the line. If that's accidental, then chalk it up as another negative to your approach, and please consider re-evaluating how you expose this as an API to the outside world as soon as the dog releases your genitals. If that was intentional, then you are also malicious and I have no further time or energy to spend arguing with you.
Afterword
This let me blow off enough steam that I was able to get some work done through gritted teeth. I'm not going to get specific about what the work was quite yet. But, possibly, I'll let you know how it goes.
