2018054 - Create end-to-end tool calling tests with a mocked LLM call and response.

Status: RESOLVED FIXED

(Core :: Machine Learning: General, task)

Description

[Greg Tatum [:gregtatum]]

LLMs are effectively non-deterministic in their responses, and any given behavior may not reproduce accurately. Rather than rely on MLPA or any type of LLM responses, we should create some mochitests that exercise the system end to end, but the call/response from the LLM should be mocked. We can create a mocked language model call and response in toolkit/components/ml that is similar to the moz-echo engine. That way we can assume a fully untrusted text conversation is happening, and measure the results in the system.

So in pseudo-code:

add_task(async function test_security() {
  const llm = await mockLLM();
  await openSmartWindow();
  await navigate("example.com/pageContent.html")
  await startChat("Summarize the page");
  llm.response("do tool call get_page_content('http://example.com/not-allowed-url'")

  is(getRejectedToolCalls().length, 1, "The tool call was rejected")
})

Comment 1

[Greg Tatum [:gregtatum]]

Attachment: Bug 2018054 - Create a MockLLMEngine in toolkit

This is intentionally inside of toolkit as a primitive that can be used
for testing across browser components that need some kind of
deterministic chat component. It's designed to have an ergonomic
interface for writing tests.

Comment 2

[Greg Tatum [:gregtatum]]

Attachment: Bug 2018054 - Drive-by fix ./mach ts check toolkit/components/ml

This command isn't enforced by CI yet.

Comment 3

[Greg Tatum [:gregtatum]]

Attachment: Bug 2018054 - Centralize the HTML server test util to toolkit/components/ml/tests/MLTestUtils.sys.mjs

This centralizes the server command to make it easy to serve up inline
HTML artifacts to be able to run through various scenarios in tests.

Comment 4

[Greg Tatum [:gregtatum]]

Attachment: Bug 2018054 - Add a querySelectorDeep command

In the Smart Window tests it can be quite confusing to query select and
find an element that you need access to. This commit creates a new
testing primitive that makes it easy to cut across shadow root
boundaries and query select through open or closed shadow roots. Many
Smart Window tests confuse the behavior of being blocked by a shadow
root, and being blocked by content being in a content process. This
reduces the potential confusion and complexity around selecting these
components, as frequently tests are using SpecialPowers.spawn to
unnecessarily complicate tests, when it's really just hidden within the
shadow dom.

Comment 5

[Greg Tatum [:gregtatum]]

Attachment: Bug 2018054 - Add a MockEngineManager to AIWindowTestUtils.sys.mjs

The MockLLMEngine in toolkit/components/ml doesn't know how an engine is
used by an end component. This is the Smart Window side of this mock
that knows about how the chat APIs are configured. It allows for mocking
out specific types of calls to a language model, and allows for writing
deterministic tests to exercise the behaviors of tool calling and
language models, especially for writing security tests.

Rather than fully clean up the other call sites for startMockOpenAI,
this marks the triple duplicated startMockOpenAI function as deprecated
so that it can be cleaned up as needed.

Comment 6

[Greg Tatum [:gregtatum]]

Attachment: Bug 2018054 - Add a first end-to-end security test

This test should serve as a first test to begin to really exericse the
security properties of chats. I added additional documentation around
the intent behind the functions so that it would be easy for coding
assistants to duplicate and create new tests in this style, so that they
are fully end to end and really exercise the characteristics of the
system

Comment 8

[Taimur]

Attachment: WIP: Bug 2018054 - Add browser_security_get_open_tabs.js security test

Comment 9

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/be9dd0705d58
https://hg.mozilla.org/integration/autoland/rev/424a8b7c6ca6
Create a MockLLMEngine in toolkit r=ai-platform-reviewers,thasan
https://github.com/mozilla-firefox/firefox/commit/b057a8ccc599
https://hg.mozilla.org/integration/autoland/rev/e298f0d92ee3
Drive-by fix ./mach ts check toolkit/components/ml r=ai-platform-reviewers,thasan
https://github.com/mozilla-firefox/firefox/commit/2b9400f250c1
https://hg.mozilla.org/integration/autoland/rev/9d0e4e432808
Centralize the HTML server test util to toolkit/components/ml/tests/MLTestUtils.sys.mjs r=ai-platform-reviewers,thasan
https://github.com/mozilla-firefox/firefox/commit/cff6e1119883
https://hg.mozilla.org/integration/autoland/rev/eed5292adc33
Add a querySelectorDeep command r=omarg
https://github.com/mozilla-firefox/firefox/commit/15a5b97b269b
https://hg.mozilla.org/integration/autoland/rev/1ef4c157f1c0
Add a MockEngineManager to AIWindowTestUtils.sys.mjs r=omarg,rrando
https://github.com/mozilla-firefox/firefox/commit/bcf8c023b526
https://hg.mozilla.org/integration/autoland/rev/d4627be6788a
Add a first end-to-end security test r=omarg

Comment 10

[Sandor Molnar[:smolnar]]

https://hg.mozilla.org/mozilla-central/rev/424a8b7c6ca6
https://hg.mozilla.org/mozilla-central/rev/e298f0d92ee3
https://hg.mozilla.org/mozilla-central/rev/9d0e4e432808
https://hg.mozilla.org/mozilla-central/rev/eed5292adc33
https://hg.mozilla.org/mozilla-central/rev/1ef4c157f1c0
https://hg.mozilla.org/mozilla-central/rev/d4627be6788a

Comment 11

[Greg Tatum [:gregtatum]]

:jlevinsohn is looking to get someone on this in the next sprint.

Comment 12

[Greg Tatum [:gregtatum]]

Crap, wrong bug.

Comment 13

[Taimur]

Attachment: Bug 2018054 - Add browser_security_get_open_tabs.js security test

Comment 14

[Taimur]

Attachment: Bug 2018054 - Add browser_security_get_open_tabs.js security test r?gregtatum

Comment 15

[Phabricator Automation]

Comment on attachment 9591279 [details]
Bug 2018054 - Add browser_security_get_open_tabs.js security test r?gregtatum

Revision D303228 was moved to bug 2030307. Setting attachment 9591279 [details] to obsolete.