Thanks — great question. I’m responding to your ask: “Is there any way in LiveKit to access the internal ‘thoughts’ (intermediate reasoning) of an LLM?” Short answer: Yes — LiveKit supports accessing or intercepting an LLM’s internal reasoning in scenarios where the model exposes those intermediate steps, and LiveKit provides hooks to include, strip, or modify that content before it’s forwarded (for example, to transcripts or TTS). What you can do right away • Use a model/endpoint that surfaces chain-of-thought or "thinking" output (for example, certain Realtime LLMs that support a thinking stream). • Configure the model’s thinking behavior via the model client (e.g., set

thinking_config.include_thoughts=True

) so the model forwards its internal reasoning. • Use LiveKit’s LLM output replacement/processing hooks (a custom

llm_node

or output-replacement pipeline) to intercept the thinking tokens or tags, then: ◦ Keep them in transcripts, ◦ Strip or filter them before sending to TTS, or ◦ Transform them into a different format for logging/inspection. When you might want to strip or modify thoughts • If you want natural-sounding spoken output, you’ll typically remove internal tags or reasoning markers before TTS. • If you need to audit or debug model behavior, keep the thoughts in logs or a developer-only transcript. Quick implementation pointers 1. If you’re using a realtime model with a thinking option, enable it via the model client (example:

thinking_config.include_thoughts=True

). 2. Add a processing node that inspects model outputs and handles special tags (examples are tags like

<tool_call>

or text patterns used by some models). 3. Decide the destination: ◦ Persisted logs/transcripts for debugging or analytics. ◦ Filtered text for TTS to avoid exposing internal reasoning to end users. ◦ Full thoughts for developer tools or moderated UIs. A couple of clarifying questions so I can give exact steps or sample code • Which LLM and integration are you using (e.g., Gemini Realtime API, Groq/DeepSeek, another provider)? • Do you want thoughts visible to end users, only to developers/admins, or only for internal logging? • Are you planning to surface thoughts in text only, or also in audio (TTS)? Tell me those details and I’ll provide a concise example (code snippet and where to plug it into LiveKit) tailored to your setup. Sources: LLM Output Replacement | LiveKit Documentation | Modifying LLM output before TTS | LiveKit Documentation