Skip to main content

✨ Enterprise Features - End-user Opt-out, Content Mod

Features here are behind a commercial license in our /enterprise folder. See Code

Features:

  • Content Moderation with LlamaGuard
  • Content Moderation with Google Text Moderations
  • Content Moderation with LLM Guard
  • Reject calls from Blocked User list
  • Reject calls (incoming / outgoing) with Banned Keywords (e.g. competitors)
  • Tracking Spend for Custom Tags

Content Moderation with LlamaGuard

Currently works with Sagemaker's LlamaGuard endpoint.

How to enable this in your config.yaml:

litellm_settings:
callbacks: ["llamaguard_moderations"]
llamaguard_model_name: "sagemaker/jumpstart-dft-meta-textgeneration-llama-guard-7b"

Make sure you have the relevant keys in your environment, eg.:

os.environ["AWS_ACCESS_KEY_ID"] = ""
os.environ["AWS_SECRET_ACCESS_KEY"] = ""
os.environ["AWS_REGION_NAME"] = ""

Customize LlamaGuard prompt

To modify the unsafe categories llama guard evaluates against, just create your own version of this category list

Point your proxy to it

callbacks: ["llamaguard_moderations"]
llamaguard_model_name: "sagemaker/jumpstart-dft-meta-textgeneration-llama-guard-7b"
llamaguard_unsafe_content_categories: /path/to/llamaguard_prompt.txt

Content Moderation with LLM Guard

Set the LLM Guard API Base in your environment

LLM_GUARD_API_BASE = "http://0.0.0.0:8000"

Add llmguard_moderations as a callback

litellm_settings:
callbacks: ["llmguard_moderations"]

Now you can easily test it

  • Make a regular /chat/completion call

  • Check your proxy logs for any statement with LLM Guard:

Expected results:

LLM Guard: Received response - {"sanitized_prompt": "hello world", "is_valid": true, "scanners": { "Regex": 0.0 }}

Content Moderation with Google Text Moderation

Requires your GOOGLE_APPLICATION_CREDENTIALS to be set in your .env (same as VertexAI).

How to enable this in your config.yaml:

litellm_settings:
callbacks: ["google_text_moderation"]

Set custom confidence thresholds

Google Moderations checks the test against several categories. Source

Set global default confidence threshold

By default this is set to 0.8. But you can override this in your config.yaml.

litellm_settings: 
google_moderation_confidence_threshold: 0.4

Set category-specific confidence threshold

Set a category specific confidence threshold in your config.yaml. If none set, the global default will be used.

litellm_settings: 
toxic_confidence_threshold: 0.1

Here are the category specific values:

CategorySetting
"toxic"toxic_confidence_threshold: 0.1
"insult"insult_confidence_threshold: 0.1
"profanity"profanity_confidence_threshold: 0.1
"derogatory"derogatory_confidence_threshold: 0.1
"sexual"sexual_confidence_threshold: 0.1
"death_harm_and_tragedy"death_harm_and_tragedy_threshold: 0.1
"violent"violent_threshold: 0.1
"firearms_and_weapons"firearms_and_weapons_threshold: 0.1
"public_safety"public_safety_threshold: 0.1
"health"health_threshold: 0.1
"religion_and_belief"religion_and_belief_threshold: 0.1
"illicit_drugs"illicit_drugs_threshold: 0.1
"war_and_conflict"war_and_conflict_threshold: 0.1
"politics"politics_threshold: 0.1
"finance"finance_threshold: 0.1
"legal"legal_threshold: 0.1

Enable Blocked User Lists

If any call is made to proxy with this user id, it'll be rejected - use this if you want to let users opt-out of ai features

litellm_settings: 
callbacks: ["blocked_user_check"]
blocked_user_id_list: ["user_id_1", "user_id_2", ...] # can also be a .txt filepath e.g. `/relative/path/blocked_list.txt`

How to test

curl --location 'http://0.0.0.0:8000/chat/completions' \
--header 'Content-Type: application/json' \
--data ' {
"model": "gpt-3.5-turbo",
"messages": [
{
"role": "user",
"content": "what llm are you"
}
],
"user_id": "user_id_1" # this is also an openai supported param
}
'

Using via API

Block all calls for a user id

curl -X POST "http://0.0.0.0:8000/user/block" \
-H "Authorization: Bearer sk-1234" \
-D '{
"user_ids": [<user_id>, ...]
}'

Unblock calls for a user id

curl -X POST "http://0.0.0.0:8000/user/unblock" \
-H "Authorization: Bearer sk-1234" \
-D '{
"user_ids": [<user_id>, ...]
}'

Enable Banned Keywords List

litellm_settings: 
callbacks: ["banned_keywords"]
banned_keywords_list: ["hello"] # can also be a .txt file - e.g.: `/relative/path/keywords.txt`

Test this

curl --location 'http://0.0.0.0:8000/chat/completions' \
--header 'Content-Type: application/json' \
--data ' {
"model": "gpt-3.5-turbo",
"messages": [
{
"role": "user",
"content": "Hello world!"
}
]
}
'

Tracking Spend for Custom Tags

Requirements:

  • Virtual Keys & a database should be set up, see virtual keys

Usage - /chat/completions requests with request tags

Set extra_body={"metadata": { }} to metadata you want to pass

import openai
client = openai.OpenAI(
api_key="anything",
base_url="http://0.0.0.0:8000"
)

# request sent to model set on litellm proxy, `litellm --model`
response = client.chat.completions.create(
model="gpt-3.5-turbo",
messages = [
{
"role": "user",
"content": "this is a test request, write a short poem"
}
],
extra_body={
"metadata": {
"tags": ["model-anthropic-claude-v2.1", "app-ishaan-prod"]
}
}
)

print(response)

Viewing Spend per tag

/spend/tags Request Format

curl -X GET "http://0.0.0.0:4000/spend/tags" \
-H "Authorization: Bearer sk-1234"

/spend/tagsResponse Format

[
{
"individual_request_tag": "model-anthropic-claude-v2.1",
"log_count": 6,
"total_spend": 0.000672
},
{
"individual_request_tag": "app-ishaan-local",
"log_count": 4,
"total_spend": 0.000448
},
{
"individual_request_tag": "app-ishaan-prod",
"log_count": 2,
"total_spend": 0.000224
}
]