FEAT Prompt Shield #271
FEAT Prompt Shield #271
Conversation
|
@microsoft-github-policy-service agree company="Microsoft" |
…and scorer so they can be imported.
… two of the fields (userPrompt, documents) of the Prompt Shield endpoint. Expanding on the docs
…re the source is an ArXiv paper.
|
Great work @ValbuenaVC ! I think you can remove the "DRAFT" tag from the title as this PR is actively being reviewed. |
Format imports to standard Co-authored-by: Raja Sekhar Rao Dheekonda <43563047+rdheekonda@users.noreply.github.com>
…ith the rest of PyRIT.
…naVC/PyRIT into t-vivalbuena/PromptShield Merging to combine suggested changes with local changes.
…naVC/PyRIT into t-vivalbuena/PromptShield Adding jupytext-generated .py files for the Prompt Shield tutorials.
| "name": "stderr", | ||
| "output_type": "stream", | ||
| "text": [ | ||
| "Error during table creation: (duckdb.duckdb.IOException) IO Error: File is already open in \n", |
There was a problem hiding this comment.
I recommend re-running without error
There was a problem hiding this comment.
(this can happen if you have your db open e.g. in dbeaver)
| load_default_env() | ||
|
|
||
| # + | ||
| #NOTE: This is throwing an IOError, but I'm not sure why? |
| ) | ||
|
|
||
|
|
||
| # with ScoringOrchestrator( |
| { | ||
| "cell_type": "code", | ||
| "execution_count": null, | ||
| "metadata": {}, |
| @@ -30,6 +30,7 @@ def __init__( | |||
| self, | |||
| prompt_target: PromptTarget, | |||
| prompt_converters: Optional[list[PromptConverter]] = None, | |||
| normalizer: Optional[PromptNormalizer] = None, #NOTE Added | |||
There was a problem hiding this comment.
Do we need this new parameter? Can we get rid of it?
If we wanted to inject this for tests, I could see it being useful, but as far as I can see it's not currently used?
| ) -> list[PromptRequestResponse]: | ||
| """ | ||
| Sends the prompts to the prompt target. | ||
| """ | ||
|
|
||
| # This happens often when trying to send one prompt, but forgetting to wrap it in brackets. | ||
| if not isinstance(prompt_list, list): |
There was a problem hiding this comment.
In the case a user passes a string, should we treat it like a list of one?
| self._memory = memory if memory else DuckDBMemory() | ||
| self._target: PromptShieldTarget = target | ||
|
|
||
| async def score_async(self, request_response: PromptRequestPiece | PromptMemoryEntry) -> list[Score]: |
There was a problem hiding this comment.
we updated this interface to add one extra field; but it can throw an error if passed
|
|
||
| def __init__( | ||
| self, | ||
| target: PromptShieldTarget, |
There was a problem hiding this comment.
nit: can we rename to prompt_shield_target so we can keep in mind it's specific?
| self, | ||
| request_response: Any | ||
| ) -> None: | ||
| if not isinstance(request_response, PromptRequestPiece) and not isinstance(request_response, PromptMemoryEntry): |
There was a problem hiding this comment.
we prob don't need to do type checking, but it may be nice to only check text due to the target only supporting text
There was a problem hiding this comment.
(mypy should take care of type checking more or less)
| ) -> None: | ||
|
|
||
| self.scorer_type = "true_false" | ||
| self._conversation_id = str(uuid.uuid4()) |
There was a problem hiding this comment.
this should be created in score_async so previous conversations aren't sent
| def sample_conversations() -> list[PromptRequestPiece]: | ||
| return get_sample_conversations() | ||
|
|
||
| @pytest.fixture |
There was a problem hiding this comment.
Can you check code coverage percentages and comment here on them?
Description
Added:
Tests
TODO