Assistant does not respect lens when searching

double

I created a lens to search only a specific site:

Then, I created a custom assistant using that lens:

And the Assistant does not respect the lens settings:

Here is the thread: https://kagi.com/assistant/6219b935-355d-4e44-be3c-fb3d21b8ea86

The Assistant should have used the Lens when searching for an answer.

tschoesi

Same for me

Benjamint22

I've been experiencing the same issue when trying to force the assistant to only search on Reddit. It picks up a small number of Reddit results, but mostly other sites. Here's my thread: https://kagi.com/assistant/dc40b5ed-fbf2-47a8-b643-c85e02bcfbb0

I tried many things to fix this and none worked, including:

Reducing the number of sites in my lens.
Selecting a different assistant then selecting my custom assistant again.
Removing my custom assistant's custom prompt.
Prepending my custom assistant's custom prompt with "Search only on Reddit."
Changing the model (I tried Claude 3 Opus and GPT 4o)
Unchecking "Personalized results"
Changing the custom assistant's lens to the default "Forums" lens
- Unfortunately, it still searched on non-forum sites
Unchecking "Web Access" in the /assistant page
- At this point, it pretended to be using results from the internet but didn't actually cite anything

Here's what my custom prompt looks like:

Search only on Reddit.

List each result/thread separately in the format following format, nesting all of their recommendations underneath each of them. Consider placeholders in square brackets optional and placeholders in parentheses required:
(Result number). (Thread title)
   - Recommendations:
     1. [DETAILED] & [PROFESSION: ...] & [CITED] & [ANECDOTAL] & [SHORT] & [POSITIVE/NEUTRAL/NEGATIVE]: (Recommendation 1)
     2. ...etc
For each recommendation, whether it's a product or simply a tip, mention whether:

they are detailed

they are made by a professional

do not consider them to be made by a professional if they don't explicitly state or imply their profession.

do not consider admin nor mod to be a profession.

they cite external sources

they are purely anecdotal/personal

they are short/brief

etc.

Once all results have been listed, make a final summary highlighting the best brands, models, and kinds based on how often they were mentioned and how recommended they are. Give the most weight to detailed or professional advice

edapstah

I'm experiencing the same issue. It's a bit disheartening, I upgraded to Ultimate solely because of Lense-enhanced assistants.

In my case, the assistant did seem to respect Lenses for some queries, but not for others. It also seemed sometimes to completely ignore all proprompting or instructions, as if the query were processed by the base un-instructed model. In some cases it would just switch to other models; This happened often when regenerating responses, it'll stop using the custom assistant's designated model, and find its way to Claude.

*Edit: For reference, I have a lense set to only *.msdmanuals.com/professional, *.merckmanuals.com/professional, *.fpnotebook.com
(Thought it might be an issue that I included /professional initially, but it made no difference)

Quite unusable for me in its current state.

bfulop

I have exactly the same issue.

I have set up a Lens that searches the documentation of a TypeScript library: https://kagi.com/lenses/5yh1EkL9OyHV7OQdB9ZqjuBSpddt77ji

I've set up an assistant to use this lens.
Here's the answer it gave for a very specific question about a combinator from this library:
https://kagi.com/assistant/9309e41b-c0cd-4e3e-992e-ebf89acb287c

This is very clearly totally off from the expected a response. (do a search for Context.Tag with the lens).

bfulop

It seems to be a regression. When I switched to Ultimate a couple of weeks ago it was specifically because I tested the lens feature and it was working well.

Luis

Hey all! We apologise for this issue. We're actively working on a fix.