FIX - Ollama Prompt Rendering & Course Documentation Improvements

[djscheuf]

• Fixes Ollama Wrapper to render incoming prompts that reference a file with the files contents
  • Depends on being provided a 'documentRoot' value, which is where the prompt rendering will attempt to discover the file path from. e.g. if @prompt-under-test.md is located at the root of the project, and appears as such in the prompt sent to ollama, the wrapper will attempt to find the file contents by looking at path {root of the project}/prompt-under-test.md (ONLY in Provider Mode)
• Updates Ollama Wrapper to support both Ollama and OpenAI API response standards. Should improve flexibility depending on if you are running something like Ollama or LM Studio for your local models.
• Updated Ollama Wrapper GRADER Mode to simple pre-pend the System Prompt, avoiding uneven model support for System type messages.
  • Also adds a response guard to the GRADER mode, ensuring that if the model rambles in addition to the json grader output, the specific grader output is extracted and sent, so Promptfoo doesn't fall over.
• Updated the How to Trust AI Readme
  • Added Install/Setup instructions listing steps to get setup for the class, including confirmation scripts for each step.
• Updated the existing promptfoo config (both repo demo and course config) to use the provider labels when specifying the defaultTest options for graders.
  • Promptfoo only needed the former exec ... label because the id for our wrappers was also serving as the label.
  • Now that all our providers are labelled ( supporting multiple model tests), we can use the labels to specify the provider to use for grading.
  • This further allows complete Local EDD with Ollama, as the ollama config can be used both for Evaluation and Grading.
    • Without the use of label based grader assignment, you cannot use Ollama to grade performance.

[djscheuf]

Realized I need to make another tweak to internals of Wrapper for improved stability