o3 Bad interpretation of dollar signs in the response make it illegible

[aceballos-prode]

What happened?

I found a closed issue similar to whay I am experiencing here.

When receiving a response containing dollar signs ($), such as queries for MongoDB, the system seems to interpret it as an equation making it very difficult to interpret the original (raw) response, since some dollar signs are ommited.

Version Information

v0.7.7

Steps to Reproduce

As a model to generate a response containing dollar signs, for example generating a MongoDB aggregation query such as:

"Generate a mongodb aggregation to retrieve the 20 users with the most documents"

What browsers are you seeing the problem on?

Chrome

Relevant log output

Screenshots

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

If this bothers you, I suggest turning off markdown rendering for your user messages:

[aceballos-prode]

Thank you @danny-avila. The problematic screenshoots however are not from the user messages but the model responses that, from what I have been able to see, are always formatted as markdown even with that option disabled.

Is there a similar feature to see the raw/unformatted model response? I just saw it is possible to see it by "editing" it, which is fine as a workaround, but it's not as clean as the solution you have provided for the user message.

[danny-avila]

in this case, some additional prompting is required. Most models are pretty good about correct LaTeX according to markdown standards, which models are you using? something in your instructions (preferably using an agent) about writing LaTeX in markdown format will help

[aceballos-prode]

I am not following why that would be helpful in this case. In my prompts I am asking the models (mostly OpenAI o3-mini) to generate a MongoDB aggregation query. The model does generate it correctly and, as it should, includes multiple dollar signs.

Since it contains dollar signs but it is not LaTeX, it is formatted incorrectly as an equation, preventing the user from seeing the correct dollar signs, as shown in the example below.

Model raw reponse: correct MongoDB query with required dollar signs, as seen in the "edit response" view:

[ { $group: { _id: "$userId", count: { $sum: 1 } } }, { $sort: { count: -1 } }, { $limit: 15 } ]

Model output visible to the user, some dollar signs missing due to incorrect interpretation. If copied directly from the interface the query is not functional due to missing dollar signs:

[ { group: { _id: "userId", count: { sum: 1 } } }, { sort: { count: -1 } }, { $limit: 15 } ]

Below there are two screenshots of the examples above.

For my needs, I can access the correct code via the copy icon or entering the edit view, but in the long run I thought this would be something worth looking into, as dollar signs are relatively common in programming and scripting and I can easily see how this behavior would replicate in other cases.

[danny-avila]

Now that you said o3, i believe this is the issue you are having:

#5601 (comment)

[danny-avila]

In LC, you can set the custom instructions to "Formatting re-enabled" and it should help

[aceballos-prode]

You are exactly right. I was not aware of this change with OpenAI reasoning models and didn't realize the model was replying in plain text. Thanks!