Why does the LLM sometimes lost context when it ge...
# ask-ail
lively-coat-10905
08/19/2025, 11:13 AMWhy does the LLM sometimes lost context when it gets interrupted?
t
tall-airline-16756
08/19/2025, 11:13 AMThanks for reaching out, Computer. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.
tall-airline-16756
08/19/2025, 11:14 AMWhen an LLM loses context during interruptions, it’s usually because the system trims down the conversation history to keep things accurate. Basically, when there’s an interruption, the system only keeps the part of the response it actually heard before the interruption happened. This helps keep the conversation history precise, but it can mean the LLM loses track of the full response it was working on.
Sometimes, timing issues between the audio playback and text processing can cause the system to cut off at the wrong spot, which adds to the context getting lost.
If you’d like to tweak how interruptions work, you can play around with settings like:
•
allow_interruptionsFalseinterrupt_speech_durationinterrupt_min_wordsl
lively-coat-10905
08/19/2025, 11:20 AMThe LLM asked me a question and when I answered, it though the answer was for the previous question and asked the question again. It's almost like the LLM output I interrupted was not added to the context before generating the next reply
t
tall-airline-16756
08/19/2025, 11:22 AMHey there! The behavior you're noticing—where the LLM repeats questions after interruptions—happens because sometimes the interrupted output doesn’t get added to the conversation context. Specifically, if the interruption happens before the agent starts speaking, the system doesn’t save that interrupted response. That’s why the LLM loses track of what it was about to say and ends up treating your next input like it’s responding to the previous question.
Here’s a quick rundown of how interruptions are handled:
• If the agent is interrupted while it’s speaking, the system cuts off the response and adds it to the context with an
interrupted=Trueinterrupt_speech_durationl
lively-coat-10905
08/19/2025, 11:25 AMBut I interrupted after the agent had already spoken a full sentence. But it still lost context
t
tall-airline-16756
08/19/2025, 11:25 AMHey there! Just a quick heads-up: even if the agent finishes a full sentence before being interrupted, the system only keeps the part of the response the user actually heard up to that interruption. So, if the interruption happens mid-playback, only that portion stays in the conversation history, and the rest gets dropped. This way, the chat history truly reflects what both sides experienced.
We designed it this way to keep things accurate, but if interruptions happen a lot or at tricky moments, some context might get lost. To help with that, you can:
• Tweak settings like
interrupt_speech_durationinterrupt_min_words