268 enable configuration of the llm evalutation defence in sandbox

[pmarsh-scottlogic]

completes #268

changes

This PR works to make the instructions for the evaluation LLM cofigurable in the sandbox. See the screenshot

• renames the defence type LLM_EVALUATION to EVALUATION_LLM_INSTRUCTIONS
• adds defence configs to the DefenceInfo objects for EVALUATION_LLM_INSTRUCTIONS
• populates prompts with default values from the templates
• renames some methods and constants to disambiguate language
  

Language diambiguation

• We have the Evaluation LLM as part of our langchain architecture
• The evaluation LLM is actually made of two separate models, which I call prompt injection eval and malicious prompt eval
• Each of these models, when initialised in langchain take a prompt template. This is like a prompt, but with placeholders which then become replaced dynamically with concrete values by langchain at runtime, to make a prompt value
• The prompt value is the concrete prompt with no placeholders that is given the the model.
• A prompt template is made up of a pre prompt prepended to a main prompt.
• The pre prompt is what is configurable by the user. It is what tell the model how to behave.
• The main prompt contains the actual question asked to the LLM, with placeholder values. See promptTemplates.ts for examples

concerns with the PR

• the text box that holds the prompt is necessarily quite big. I wonder if we should add a scroll bar past a threshold
• I wonder if the names I came up with are too much of a mouthful

a potential helpful refactor for a different ticket

• I wonder if we should refactor the Defence Config ids so that they are a shared enum or type across the front and backend rather than arbitrary strings. It didn't cause my any trouble yet but it strikes me as being vunerable to easy-to-make, hard-to-debug mistakes.

[pmarsh-scottlogic]

Happy to undo those name changes btw if y'all disagree with them. I mostly did it for my own benefit

[gsproston-scottlogic]

Looks good to me! Just a couple of things to tweak.

[chriswilty]

A few suggestions, and some other general comments on the code where we are making life a bit difficult for ourselves. I will add a few new issues to cover the latter.

One final comment - our separation of concerns in the back-end is a bit fuzzy, and it's making it difficult to see what each type's responsibility is. I'll add an issue to look into that as well.

[gsproston-scottlogic]

Nice work 👍

[chriswilty]

@pmarsh-scottlogic Pushed some minor code cleanup, and solved the jest problem. This one's ready to go.

[chriswilty]

I've updated the UI to preserve whitespace in the defence prompt boxes, which makes it more readable: