Application suddenly can't start any functionality or understand what is said

[Lovenoreus]

I had this error occour today on a commit that was previously working from 14th december, and it gives the same logging errors when pulling the latest change

It repeats the intro messages and claims it can't understand me, shown in the log bellow

It has no code changes and uses default settings like region.
log_new_commit.txt

[Lovenoreus]

When the application worked for me there were also some other major core-functionality problems that makes the application not usable in real life. Are these things that you also experience?

[Stopping] - The phone center chooses not to respond to some of the callers answers and you say your sentence again
[Sentences cutting off] - The log shows that all words are being heard, both in "partial" and "complete" yet the majority of sentences get their endings cut off when looking at the conversation from the UI and the full sentence is not sent to OpenAI.
[Sudden language change] - When using English you can answer all the bots follow-up questions but then afterwards it will change to French and get confused about what to do next
[Repetition] - The bot will say a full sentence, then say the exact same sentence again only adding an additional word

I would assume that you also have these problems every call since we both follow the same setup instructions, but you seem to be adding more advanced extra features which makes me think that core pieces work better on your end.

[clemlesne]

Language mitigation is shipped in v17.0.2.

[clemlesne]

Regarding the word missed during recognition, like mentioned in the documentation, a configuration exists: https://github.com/microsoft/call-center-ai?tab=readme-ov-file#customize-the-conversation

You can fine tune recognition_stt_complete_timeout_ms depending on your environment. More wait time will be traduced on more probability to the SST to transcribe the voice, in exchange to latency.

For example, with French, 100ms renders well. For English, the model is more performant, 50ms is OK.

[Lovenoreus]

I am always testing in English so the standard value of 100ms should be sufficient is what you are saying

[Lovenoreus]

I will write the issues up and post the voice recording next time absolutely

[clemlesne]

Yes 100ms should be fine.

Also notice that running your app locally create a lot of latency (from your antivirus stack notably, and often a bad network). It's always quicker in the cloud.

[Lovenoreus]

When calling the cloud container the call will not reach the app at all, and when turning off the container and start the local-app that is when you can reach it but these log lines appear

[clemlesne]

Seems related to a bug in the Azure Inference API, which is now mitigated: #391. I'll publish a new version in the next hours.

[clemlesne]

Tool call bug fix is shipped in v17.0.2.

[Lovenoreus]

Yes I tried with the newest commit today using 17.0.3 in the makefile and it works, along with some major functionality problems that bot stops working after a few messages. I will post this problem as well as an issue with proof. If you write down a longer (10 second) question in exact words, ask it to the bot and then review the /report you will see how much it does not retain and when it looses the ability to continue the conversation

[clemlesne]

Waiting four your reproduction steps

[clemlesne]

Versions are shipped frequently. When a breaking change is introduced, a major version is published and the container is tagged main. If you deployed your code with the same tag, you'll use that version which contains a breaking change.

To avoid that, hardcode the version you are deploying, like advised in the installation steps: https://github.com/microsoft/call-center-ai?tab=readme-ov-file#3-run-deployment-automation.

[Lovenoreus]

I understand now how it works, I don't see why it would make local stop working (unless the config.yaml would have been altered radically)

But when I redployed it with the newest image_version specifically and commit it worked

[clemlesne]

Indeed the config format changed a bit from v16 to v17. This explains the major version number change.