Streaming Responses

Back to Prompts Library

Streaming responses enable real-time, incremental LLM output—the model transmits output pieces as they become available instead of waiting for a complete response. This reduces latency and improves interaction for conversational AI, speech-to-text, and content generation tools.

---

Current Capabilities

• Voice: Supports the Voice Gateway channel for GenAI features including Agent Node, with Deepgram as the supported TTS engine.
• Chat: Delivers real-time LLM response streaming for Web/Mobile SDK channels. Agent Node and Prompt Node stream responses token-by-token.
• Models: Integrates with OpenAI and Azure OpenAI out of the box. Custom prompts support other streaming-capable LLMs.
• Node-level activation: Streaming activates when a streaming prompt is selected at the node level, even if the feature-level prompt doesn’t use streaming.

Note: Agent Node Tool Calling and Streaming

• V1 Custom JavaScript Prompts: Supports tool calling and streaming as separate capabilities only—not simultaneously.
• V2 Custom JavaScript Prompts: Supports both tool calling and streaming together using the OpenAI/Azure OpenAI response format.

Benefits

Benefit	Description
Real-Time Output	Text appears instantly, reducing wait times.
Lower Latency	Faster response time improves user experience.
Improved Interaction	Partial outputs support iterative writing and brainstorming.
Live Application Support	Enhances real-time chat, speech-to-text, and code autocompletion.

---

Use Cases

Domain	Use Cases
Healthcare	Streaming patient history summaries; real-time clinical study analysis.
Finance	Streaming portfolio breakdowns; incremental compliance document summaries.
E-commerce	Streaming side-by-side product comparisons for informed decisions.
Education	Delivering detailed course outlines or study material summaries.
Legal	Streaming legal precedent explanations; incremental contract analysis.
Customer Support	Streaming detailed troubleshooting steps for complex issues.
Human Resources	Streaming HR policy or benefits explanations for employees.
Marketing	Streaming in-depth campaign analysis and ROI breakdowns.

---

Enable Streaming

Use a Default Streaming Prompt

The Platform provides default streaming prompts for Agent Node and Prompt Node for OpenAI models.

Go to Generative AI Tools > GenAI Features and select the default-streaming prompt for the required feature.

Create a Custom Streaming Prompt

See How to Add a Custom Prompt and enable the Stream Response toggle.

The streamed response must follow this format:

Field	Description
conv_status	Indicates whether the conversation has ended or is ongoing.
AI Agent response	The generated response sent to the end user.
collected entities	Stringified JSON object containing extracted entities.

• Add the required stream parameter to your custom prompt (e.g., "stream": true for OpenAI/Azure OpenAI).
• The saved prompt appears with a “stream” tag in the Prompts Library.
• Enabling streaming disables: Exit Scenario, AI Agent Response, Collected Entities, and Tool Call Request (for Agent Node).

---

Performance Benchmarks

Task	Mode	Input Tokens	Output Tokens	Time (s)	Reduction
Agent Node	Non-streaming	777	90	2.59	Output: -30%, Time: 83%
	Streaming	676	62	0.44
50-word Joke	Non-streaming	95	54	2.4	Output: +10%, Time: 80%
	Streaming	68	60	0.47
500-word Joke	Non-streaming	95	595	22.39	Output: +10%, Time: 98%
	Streaming	68	649	0.41
500-word Joke	Non-streaming	68	642	30.11	Output: -0.05%, Time: 97%
	Streaming	68	641	0.88
500-word Story	Non-streaming	68	616	16.86	Output: +2.27%, Time: 97.5%
	Streaming	68	630	0.44
500-word Essay	Non-streaming	70	687	22.23	Output: +1.46%, Time: 97.15%
	Streaming	70	697	0.63

Key insights:

• Output < 100 tokens: 80–85% time reduction.
• Output 100–600 tokens: 97–98% time reduction.
• Output > 600 tokens: 98–99% time reduction.
• Output quality variance is minimal (≤2.5%), ensuring task reliability.

These benchmarks were conducted under specific scenarios. Performance varies by environment. Conduct your own testing before enabling streaming in production.

---

Analytics

Usage Logs track and differentiate streaming and non-streaming responses.

Metric	Description
TTFT (Time to First Token)	Time until the first token appears. Blank for the final response chunk since no further messages are sent.
Response Duration	Time from the first chunk to the last chunk.
Response Type	Indicates streaming or non-streaming on the Detailed Log page.

---

Limitations

Limitation	Reason
No post-processing	Requires the complete response, which conflicts with incremental delivery.
No guardrails	Content moderation requires full-context evaluation, incompatible with token-by-token streaming.
Voice compatibility	Depends on TTS engine support for bi-directional streaming (e.g., Deepgram).
No BotKit interception	Real-time delivery is incompatible with message interception.

Streaming quality depends heavily on prompt design. LLMs are subject to hallucination—ensure prompts are accurate and aligned with desired output before using in production.