Add Parallel Search + Cartesia Line integration

[georgeatparallel]

Summary

This is a fork-only review PR before opening anything upstream.

It adds a Cartesia community integration page for Parallel Search + Cartesia Line. The example builds a Line agent with a web_search loopback tool and calls Parallel Search with mode="turbo" through the Parallel Python SDK.

Latest updates

I tightened the guide to match Cartesia's community-integration style: the quick start is step-by-step, the voice agent avoids reading citations or URLs aloud, and the example stays focused on the smallest runnable Line agent.

I also sharpened the example around the intended product message: Turbo Search inside a live voice turn. The guide calls out mode="turbo" as Parallel's lowest-latency Search mode, uses a neutral time-sensitive weather prompt, and now defaults to 3 results x 240 excerpt chars = 720 total chars with LINE_MAX_TOKENS=180 so the docs match the lower-latency demo path.

After testing the real Cartesia voice path, I added current-date prompt guidance and then moved that prompt into a per-agent system_prompt() helper so new calls get a fresh date. The agent includes date/year context for relative-time searches and tells the model to say when results are stale or conflicting instead of guessing.

Checked

• Extracted the embedded Python and compiled/imported it with cartesia-line + parallel-web
• Confirmed the documented defaults resolve to MAX_RESULTS=3, MAX_EXCERPT_CHARS=240, MAX_CHARS_TOTAL=720, and LINE_MAX_TOKENS=180
• Constructed the Cartesia LlmAgent from the extracted code and confirmed the Cerebras config path keeps reasoning_effort="low"
• Ran a live Parallel Search call with mode="turbo" and the documented defaults; it returned 3 results and formatted a voice payload successfully
• Ran the extracted docs code as a temporary Line server and checked GET /status plus Cartesia-shaped POST /chats
• Scanned the submission for stale old-default text and ran git diff --check

Note

This PR targets my fork only. It is not opened against cartesia-ai/community-integrations.

[vshulman]

are these sensible defaults?

[georgeatparallel]

seems decent - maybe not enough results? I was thinking speed first. wdyt? 5? 8?

[sergei1152]

might be a redudant objective, you can just user search queries

[sergei1152]

not sure if you want to somehow customize for voice agents in particular

[georgeatparallel]

agreed. for this search_queries=[query] is enough and keeps the SDK call easier to copy. I removed the custom objective and left voice response behavior to the agent prompt.