docs: update for pipecat PR #4423

[markbackman]

Automated documentation update for pipecat PR #4423.

Changes

New service page

• api-reference/server/services/llm/inception.mdx — Created full documentation page for InceptionLLMService
  • Documents Mercury-2 diffusion-based reasoning model
  • Includes unique parameters: reasoning_effort (instant, low, medium, high) and realtime (for reduced TTFT)
  • Notes that Inception does not support the "developer" message role
  • Installation instructions with uv add "pipecat-ai[inception]"
  • Usage examples including basic setup, custom settings, and function calling

Navigation updates

• docs.json — Added Inception to LLM services navigation (alphabetically between Groq and Mistral)
• api-reference/server/services/supported-services.mdx — Added Inception to LLM services table

Gaps identified

None

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
daily	🟢 Ready	View Preview	May 21, 2026, 4:10 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[joycech333]

Thank you for integrating support for Mercury @markbackman ! Just a note that instant mode does not support tool calls, so defaulting to medium or low reasoning effort would be best (Mercury still achieves much lower TTFT than small autoregressive models at medium)

[markbackman]

@joycech333 interesting! I found that tool calling does work with the instant setting. Here's a recording running the example. Any idea why this works with the instant setting?

CleanShot.2026-05-21.at.12.55.09.mp4

[markbackman]

@joycech333, one more question: should we default realtime to True? I can't quite tell the difference in setting this, but docs indicate that this defaults to False and is for applications requiring low TTFT (e.g. voice agents).

[joycech333]

Oh nice! To clarify, it doesn't reject tools, but we've seen more reliable tool-calling behavior at higher reasoning efforts. For a good balance between TTFT and tool-calling I think low would be a good option. And yes, default realtime to True for voice agents!

[markbackman]

Oh nice! To clarify, it doesn't reject tools, but we've seen more reliable tool-calling behavior at higher reasoning efforts. For a good balance between TTFT and tool-calling I think low would be a good option. And yes, default realtime to True for voice agents!

Good to know. I'll make that example change + default setting to the realtime setting. Here's the PR with those changes:
pipecat-ai/pipecat#4539