FEAT: Add modality support detection system for prompt targets#1381
FEAT: Add modality support detection system for prompt targets#1381fitzpr wants to merge 3 commits intoAzure:mainfrom
Conversation
fitzpr
force-pushed
the
feature/modality-support-detection
branch
from
957d4b9 to
ed5cdff
Compare
|
@fitzpr looks like now only a test file is left. ping me when the code is back 🙂 Appreciate you taking a stab at this! |
fitzpr
force-pushed
the
feature/modality-support-detection
branch
from
8a4723c to
8328b60
Compare
| target_specific_params: Optional[Dict[str, Any]] = None | ||
| """Additional target-specific parameters.""" | ||
|
|
||
| supported_input_modalities: Optional[list[str]] = None |
There was a problem hiding this comment.
@bashirpartovi may have thoughts on whether this belongs on the identifier or not
In my efforts to remove some extraneous commits i removed the files lol. Should be good now. @romanlutz |
fitzpr
force-pushed
the
feature/modality-support-detection
branch
from
b86c25b to
8328b60
Compare
| """ | ||
|
|
||
| #: OpenAI Chat targets support both text and image_path modalities | ||
| SUPPORTED_INPUT_MODALITIES = ("text", "image_path") |
There was a problem hiding this comment.
we'd probably have to do all the other targets, too
There was a problem hiding this comment.
All major targets implemented - OpenAI, TextTarget, HuggingFace targets updated with proper modality definitions
There was a problem hiding this comment.
I mean ALL 🙂 There are quite a few. I can help with that, though, if you want.
fitzpr
force-pushed
the
feature/modality-support-detection
branch
from
d2a76bd to
149b115
Compare
- Address all Roman's feedback from PR Azure#1381 - Remove quoted type annotations, add future imports - Create generic ModalityDiscovery utility class - Simplify code with direct returns and set.union - Remove confusing multimodal helper methods - Move runtime testing to reusable common module - Update tests to use new architecture - Enable any target to use discovery via protocol
fitzpr
force-pushed
the
feature/modality-support-detection
branch
from
5f23b8d to
0141366
Compare
|
|
||
|
|
||
| def verify_target_capabilities( | ||
| target: Any, |
There was a problem hiding this comment.
This should be of type PromptTarget but then we'd have a circular dependency. Maybe this should live in pyrit.prompt_target to solve that? I don't think there are downsides to this since it requires targets anyway.
| # Try the request | ||
| async def _test(): | ||
| try: | ||
| await target._async_client.chat.completions.create( |
There was a problem hiding this comment.
so this would only tell us if it works for chat completions on a openaichattarget
I would love to make this generic. For that, we should craft a proper Message for each type, matching the PromptDataType in PyRIT, and use the target's send methods so that it actually works end-to-end. Since we have that method standardized for all targets it should be possible to validate in a straightforward manner. We would need to ship one example image, audio, and video with PyRIT but we can make them tiny like you suggested. Then, we can load those and use in this function.
We also need to test combinations of modalities based on what the target says it supports.
There was a problem hiding this comment.
... and if it turns out that's not supported we need to spit out a modality configuration that can be passed to the target to update itself.
There was a problem hiding this comment.
@romanlutz The PR page seems stuck on an old state - it shows "1 commit" but we actually have 11 commits with all the set[frozenset[PromptDataType]] changes you requested.
Can you check the commits tab: https://github.com//pull/1381/commits
Our latest commit (7709e55) has exactly what you suggested! The main PR page seems to have a GitHub caching issue.
fitzpr
force-pushed
the
feature/modality-support-detection
branch
from
4358b92 to
2484292
Compare
Addresses all Roman's feedback from PR Azure#1377: - Uses set[frozenset[PromptDataType]] instead of tuples - Exact frozenset matching prevents ordering issues - Implemented across all target types (OpenAI, HuggingFace, TextTarget) - Future-proof pattern matching for new OpenAI models - Optional verification utility for runtime testing - Comprehensive test suite with 8 passing tests
The PR interface is showing cached old commits instead of our current code. This commit should trigger GitHub to display the correct current state.
2 participants
@romanlutz This addresses your feedback from PR #1377 about needing modality support detection.
This PR implements the capability detection system you suggested, allowing attacks to determine whether targets support multimodal input (text + image/video) before attempting to send multimodal messages.
Key features:
This should be merged before updating #1377 to use the new detection system.