Create a mochitest to test that URL contents are omitted when using URL tokens everywhere in the conversation
Categories
(Core :: Machine Learning: General, task, P2)
Tracking
()
People
(Reporter: gregtatum, Unassigned)
References
(Blocks 2 open bugs)
Details
(Whiteboard: [aiplatform])
Edit: Since this is a duplicate of Bug 2024354 I'm changing this to be a mochitest to assert this behavior.
Currently the entire URL from get_page_content is being passed in as a result. This has several issues. The LLM can get confused on what it's seeing when the URL is really long. It's inefficient from a cost perspective as it involves extra tokens. It's also another surface for hidden content an prompt injections. We could provide some kind of ID for the language model to use to reference something. We probably need to do some design work for the shape and schema of that. I could see keeping the host for the URL as being beneficial, and there IS signal for the path as to what the content of the link is. We can decide what that limit looks like and design a nice identifier that can contain some signal in it still.
| Reporter | ||
Updated•6 months ago
|
| Reporter | ||
Comment 1•5 months ago
|
||
Possibly will be fixed by Bug 2024354.
| Reporter | ||
Updated•5 months ago
|
Updated•5 months ago
|
| Reporter | ||
Comment 3•4 months ago
|
||
I'm converting this into a good mochitest that audits this behavior since it was fixed in Bug 2024354.
| Reporter | ||
Updated•4 months ago
|
| Reporter | ||
Updated•4 months ago
|
Description
•