vision-agents-plugins-inworld 0.6.3

Inworld AI integration for Vision Agents (LLM/VLM router, TTS, Realtime WebRTC)

Inworld AI Plugin

Inworld AI integration for Vision Agents. Provides:

• LLM / VLM — text and vision chat completions through the Inworld Realtime Router, which proxies upstream across OpenAI / Anthropic / Google / etc. with auto-selection, fallbacks, and traffic splitting.
• TTS — high-quality streaming text-to-speech.
• Realtime — WebRTC speech-to-speech for low-latency voice agents.

Installation

uv add "vision-agents[inworld]"
# or directly
uv add vision-agents-plugins-inworld

Get your API key from the Inworld Portal and set INWORLD_API_KEY in your environment (or pass api_key= explicitly).

LLM / VLM (router)

inworld.LLM and inworld.VLM hit Inworld's OpenAI-compatible /v1/chat/completions endpoint. The model argument accepts:

• "inworld/<router-id>" — a router defined in the Inworld portal
• "<provider>/<model-id>" — e.g. "openai/gpt-4o-mini"
• "auto" — let Inworld pick (combine with sort_by)

from vision_agents.plugins import inworld

# Lowest-latency routing with a small fallback chain
llm = inworld.LLM(
    model="auto",
    sort_by=["latency"],
    ttft_timeout="500ms",
    fallback_models=["openai/gpt-4o-mini", "google-ai-studio/gemini-2.5-flash"],
)

# Vision over the router (frames sent as image_url content).
# Tuned for low-latency video Q&A: small frames, short buffer, fast fallback.
vlm = inworld.VLM(
    model="auto",
    sort_by=["latency"],
    ttft_timeout="500ms",
    fallback_models=["google-ai-studio/gemini-2.5-flash", "openai/gpt-4o-mini"],
    fps=1,
    frame_buffer_seconds=3,
    frame_width=512,
    frame_height=384,
)

See example/inworld_llm_example.py and example/inworld_vlm_example.py for end-to-end voice and video agents respectively.

Tuning the VLM for latency

Video Q&A latency is dominated by input tokens (frames cost a lot more than text) and upstream choice. The example values above are the right starting point:

• frame_width=512, frame_height=384 — 4× fewer bytes than the 800×600 default, with negligible accuracy loss for typical Q&A.
• frame_buffer_seconds=3 with fps=1 — 3 frames/request. Longer buffers inflate input tokens quadratically without helping short-horizon questions.
• sort_by=["latency"] + ttft_timeout="500ms" + a fallback chain to fast vision models keeps TTFT predictable when one provider degrades.

When to use which

• Voice agents → inworld.Realtime (WebRTC, full-duplex, lowest latency). The text router cannot beat full-duplex audio for STT→LLM→TTS pipelines.
• Text agents, STT→LLM→TTS pipelines, video Q&A → inworld.LLM / inworld.VLM.

Routing kwargs

• fallback_models: ordered list, tried on failure.
• ignore_models: excluded from auto.
• sort_by: any of "price", "latency", "throughput", "intelligence", "math", "coding". Multiple metrics rank with tiebreakers.
• ttft_timeout: switch to fallback if first token doesn't arrive in time (Inworld minimum "300ms").
• metadata: free-form dict consumed by router CEL expressions for conditional routing.
• web_search / web_search_options: opt-in upstream web grounding.
• compression_aggressiveness (0–1): Inworld's prompt compression applied to the system message — cuts input tokens, lowers TTFT for long prompts.
• extra_body: raw escape hatch merged in last.

Caching

Implicit prompt caching is automatic on OpenAI / DeepSeek / Gemini-2.5 upstreams — no code needed. Explicit caching (Anthropic / Google) is a per-message thing: add a cache_control block to the message content yourself, e.g. {"type": "text", "text": "...", "cache_control": {"type": "ephemeral"}}.

Router definitions themselves (router IDs, A/B variants, traffic weights) are configured in the Inworld portal — out of scope for this plugin.

TTS

High-quality text-to-speech with streaming support. The plugin now defaults to Inworld's TTS-2 model (currently in research preview), which adds natural-language steering, 100+ languages (15 GA, 90+ experimental), and high-quality instant voice cloning over the previous inworld-tts-1.5-* generation.

from vision_agents.plugins import inworld

# Defaults to model_id="inworld-tts-2", voice_id="Sarah"
tts = inworld.TTS()

# Or specify explicitly
tts = inworld.TTS(
    api_key="your_inworld_api_key",
    voice_id="Ashley",
    model_id="inworld-tts-2",
    temperature=1.1,
)

TTS options

• api_key: Inworld AI API key (default: reads from INWORLD_API_KEY)
• voice_id: Voice to use (default: "Sarah"; "Dennis", "Ashley", "Olivia", "Clive" and custom/cloned voices also supported)
• model_id: "inworld-tts-2" (default), "inworld-tts-1.5-max", "inworld-tts-1.5-mini". "inworld-tts-1" and "inworld-tts-1-max" are deprecated by Inworld — migrate to inworld-tts-2 or inworld-tts-1.5-*.
• temperature: 0–2 (default: 1.1)

The plugin requests LINEAR16 (16-bit PCM WAV) chunks from Inworld so each streamed chunk is self-contained and decodes cleanly under streaming TTS; no extra configuration needed.

Steering (TTS-2)

TTS-2 takes natural-language stage directions inline with your text. Place the instruction in square brackets before the segment it should apply to:

text = (
    "[whisper in a hushed style] I have to tell you something. "
    "[laugh] Just kidding! [say with force] Now let's get to work."
)
async for chunk in await tts.stream_audio(text):
    ...

Steering covers articulation, intonation, volume, pitch, range, speed, and vocal style — and supports non-verbal sounds like [laugh], [breathe], [clear throat], [sigh], [cough], [yawn]. Combining dimensions ([whisper in a hushed style], [say playfully and very fast]) produces better results than bare single-word tags. See Inworld's steering docs and prompting guide for the full reference.

Agent example

A complete example wiring inworld.TTS() into a Stream-edge agent with Deepgram STT, Gemini LLM, and smart-turn detection lives at example/inworld_tts_example.py. The companion example/inworld-audio-guide.md is loaded as the agent's system prompt and teaches the LLM how to emit TTS-2 steering tags so replies sound expressive out of the box.

Realtime (WebRTC)

Low-latency speech-to-speech via Inworld's Realtime API. This transport uses WebRTC (UDP, native Opus) for lower latency than the WebSocket alternative. Requires a WebRTC-capable edge transport — pair with getstream.Edge() as shown below.

from vision_agents.core import Agent, User
from vision_agents.plugins import getstream, inworld, smart_turn

agent = Agent(
    edge=getstream.Edge(),
    agent_user=User(name="My Agent", id="agent"),
    llm=inworld.Realtime(
        model="openai/gpt-4o-mini",
        voice="Dennis",
        instructions="You are a friendly voice assistant.",
    ),
    turn_detection=smart_turn.TurnDetection(),
)

Realtime options

• model: provider-prefixed model ID. Examples: "openai/gpt-4o-mini" (default), "google-ai-studio/gemini-2.5-flash", "inworld/<router-id>" for an Inworld router
• voice: voice for audio responses (default: "Dennis"; "Clive", "Olivia" and custom voices also supported)
• api_key: Inworld AI API key (default: reads from INWORLD_API_KEY)
• instructions: system prompt
• realtime_session: advanced — pass a full RealtimeSessionCreateRequestParam for session fields not exposed by the primary args (custom turn-detection, tool_choice, etc.)

Registering tools

realtime = inworld.Realtime()

@realtime.register_function(description="Get the current weather for a city.")
async def get_weather(city: str) -> str:
    return f"It's sunny in {city}."

Tools follow the OpenAI function-calling schema. Inworld's Realtime API is protocol-compatible with OpenAI's Realtime API, so registered functions flow through the same response.function_call_arguments.done path.

Notes

• v1 is WebRTC only; a WebSocket transport may be added later.
• Video input is not currently supported by Inworld's Realtime API.

Requirements

• Python 3.10+
• httpx>=0.28, av>=10, aiortc>=1.9, openai[realtime]>=2.26,<3