Add WildGuard Guardrail Microservice

[daniel-de-leon-user293]

Description

Add WildGuard to guard against privacy, misinformation, harmful language and malicious use on user input prompts and/or output responses generated by LLMs.

Issues

n/a

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds new functionality)
• Breaking change (fix or feature that would break existing design and interface)
• Others (enhancement, documentation, validation, etc.)

Dependencies

n/a

Tests

• Successfully ran:
  • guardrails_tgi.py microservice script directly without container
  • with docker run CLI
  • with docker compose

[letonghan]

Hi @daniel-de-leon-user293 , please add the test script for WildGuard in test/guardrails folder like this.
Name the test script as test_guardrails_wild_guard_langchain_on_intel_hpu.sh since you use the hpu version of tgi.
Thanks : )

[qgao007]

Hi @daniel-de-leon-user293 ,
Ebi pointed it out to me earlier for my PR, mentioning that the following needs to be added
Also the new Dockerfile introduced by the PR needs to have an entry at:
.github/workflows/docker/compose/guardrails-compose-cd.yaml

[ashahba]

Thanks @daniel-de-leon-user293 for this PR!
I only have a few minor change requests.

[ashahba]

LGTM!

[dcmiddle]

@daniel-de-leon-user293 Consider adding documentation or commit message explaining why someone would use wildguard instead of llama guard.

@mkbhanda does OPEA have criteria for when it will accept similar features?
It's nice to have a variety of options if you know what you are looking for but having to choose among options also detracts from ease of use.

[daniel-de-leon-user293]

Thank you for your suggestion @dcmiddle. The latest commit adds a bit more description of wild guard in the README. To answer your question, we’re hoping to provide a diverse set of safety models that users can pick from.

[dcmiddle]

Thank you for your suggestion @dcmiddle. The latest commit adds a bit more description of wild guard in the README. To answer your question, we’re hoping to provide a diverse set of safety models that users can pick from.

cool. so can this be used in conjunction with llama guard? From the description you added it looks like a complementary list of topics.

[daniel-de-leon-user293]

No, it can be used in place of Llama Guard. The lists in the README are risk taxonomies that each model was trained to identify according to their respective datasets. Although similar, the models provide different classification performance for difference use cases.

If a user wanted to design an ensemble of guardrails, however, then in that case, these two models could be used in conjunction.

[sunstonesecure-robert]

It's nice to have a variety of options if you know what you are looking for but having to choose among options also detracts from ease of use.

I'd assert it's essential to have a variety of options AND you need to know what you are looking for if you are building LLM apps where security+privacy+bias detection > ease of use