Le Chat: Custom MCP Connectors, Memories

I recently upgraded a large portion of my pipeline from gpt-4.1-mini to gpt-5-mini. The performance was horrible - after some research I decided to move everything to mistral-medium-0525.

Same price, but dramatically better results, way more reliable, and 10x faster. The only downside is when it does fail, it seems to fail much harder. Where gpt-5-mini would disregard the formatting in the prompt 70% of the time, mistral-medium follows it 99% of the time, but the other 1% of the time inserts random characters (for whatever reason, normally backticks... which then causes it's own formatting issues).

Still, very happy with Mistral so far!

It is such a common pattern for LLMs to surround generated JSON with ```json … ``` that I check for this at the application level and fix it. Ten years ago I would do the same sort of sanity checks on formatting when I used LSTMs to generate synthetic data.

Some LLM APIs let you give a schema or regex for the answer. I think it works because LLMs give a probability for every possible next token, and you can filter that list by what the schema/regex allows next.

Interestingly, that gives a different response distribution from simply regenerating while the output doesn't match the schema.

It sounds like they are describing a regex filter being applied to the model's beam search. LLMs generate the most probable words, but they are frequently tracking several candidate phrases at a time and revising their combined probability. It lets them self correct if a high probability word leads to a low probability phrase.

I think they are saying that if highest probability phrase fails the regex, the LLM is able to substitute the next most likely candidate.