GLChat Events & Streaming

This guide is for external pipeline implementation purpose

Overview

This section explains how GLChat's external pipeline feature works under the hood, and how to implement real-time events that enhance user experience.

What You'll Learn

• How GLChat communicates with external pipelines

• Why OpenAI event format is required

• How to implement thinking events (showing AI reasoning)

• How to implement activity events (showing tool actions)

• Complete event streaming architecture

---

How External Pipeline Works

The Request Flow

Key Insight: Why OpenAI Format?

GLChat uses the OpenAI LM Invoker internally to process ALL external pipeline responses.

What this means:

• GLChat doesn't know or care which LLM you use (OpenAI, Anthropic, Google, etc.)

• GLChat expects events in OpenAI's native event format

• GLChat validates, parses, and renders events using OpenAI's schema

• If you send custom event formats, they won't display correctly

⚠️ Notes: Regardless of which LLM you use, always convert your events to OpenAI format before sending to GLChat.

Why This Matters

Wrong Approach ❌

# Your service emits Anthropic-format events
yield "event: thinking\ndata: {content: 'Let me think...'}\n\n"

Result: GLChat's OpenAI LM Invoker doesn't recognize this → Events not displayed

Correct Approach ✅

# Your service converts to OpenAI format
yield "event: response.reasoning_summary_text.delta\ndata: {...}\n\n"

Result: GLChat recognizes and displays the event

---

Understanding Event Types

GLChat supports three main event types that enhance user experience:

No	Event Type	What It Is	When to Use
1	Response Events (Text Output)	The actual text response from the LLM, streamed in chunks.	Always — this is your primary output.
2	Thinking Events (AI Reasoning)	Shows the AI's internal reasoning process. Available with thinking-enabled models.	When you want to show multi-step reasoning, planning, or analysis.
3	Activity Events (Tool Actions)	Shows discrete actions the AI performs, like web search.	When your LLM uses tools or external services.

---

Complete Event Examples

Let's walk through a complete conversation with all event types.

Scenario: User asks about recent AI developments with web search

User Query: "Explain to me how LLM works. Find on internet to fetch more details."

---

Step 1: Response Initialization

event: response.created
data: {"type": "response.created", "response": {"id": "resp_123", "status": "in_progress", ...}}

event: response.output_item.added
data: {"type": "response.output_item.added", "item": {"id": "msg_456", "type": "message", ...}}

event: response.content_part.added
data: {"type": "response.content_part.added", "part": {"type": "output_text", "text": ""}}

---

Step 2: Thinking Starts

Your LLM (with thinking enabled) begins reasoning.

Internal LLM Event:

# GLLM SDK EventEmitter captures this
{
    "type": "data",
    "value": {
        "data_type": "thinking_start",
        "id": "think-uuid-123",
        "data_value": ""
    }
}

You Convert to OpenAI Format:

{
    "type": "response.reasoning_summary_part.added",
    "item_id": "think-uuid-123",
    "output_index": 0,
    "summary_index": 0,
    "sequence_number": 1,
    "part": {"type": "summary_text", "text": ""}
}

SSE Event You Send:

event: response.reasoning_summary_part.added
data: {"type": "response.reasoning_summary_part.added", ...}

What happen: Thinking start event is sent to GLChat

---

Step 3: Thinking Content (Reasoning)

Your LLM shows its reasoning process.

Internal LLM Event:

{
    "type": "data",
    "value": {
        "data_type": "thinking",
        "id": "think-uuid-123",
        "data_value": "The user wants to understand how LLMs work..."
    }
}

You Convert to OpenAI Format:

{
    "type": "response.reasoning_summary_text.delta",
    "item_id": "think-uuid-123",
    "output_index": 0,
    "summary_index": 0,
    "sequence_number": 2,
    "delta": "The user wants to understand how LLMs work..."
}

SSE Event You Send:

event: response.reasoning_summary_text.delta
data: {"type": "response.reasoning_summary_text.delta", ...}

What happens: Thinking event is sent to GLChat

---

Step 4: Thinking End

Your LLM end reasoning process.

Internal LLM Event:

{
    "type": "data",
    "value": {
        "data_type": "thinking_end",
        "id": "think-uuid-123",
        "data_value": ""
    }
}

You Convert to OpenAI Format:

{
    "type": "response.reasoning_summary_text.delta",
    "item_id": "think-uuid-123",
    "output_index": 0,
    "summary_index": 0,
    "sequence_number": 3,
    "part": {"type": "summary_text", "text": ""}
}

SSE Event You Send:

event: response.reasoning_summary_part.done
data: {"type": "response.reasoning_summary_part.done ...}

What happens: Thinking end event is sent to GLChat

What user sees: Full thinking process on UI

---

Step 5: Activity Event (Web Search)

Your LLM performs web search.

Internal LLM Event:

{
    "type": "data",
    "value": {
        "data_type": "activity",
        "id": "search-uuid-456",
        "data_value": {
            "type": "search",
            "query": "Attention is All You Need Vaswani 2017 arXiv paper"
        }
    }
}

You Convert to OpenAI Format:

{
    "type": "response.output_item.done",
    "item": {
        "id": "search-uuid-456",
        "type": "web_search_call",
        "status": "completed",
        "action": {
            "type": "search",
            "query": "Attention is All You Need Vaswani 2017 arXiv paper"
        }
    },
    "output_index": 0,
    "sequence_number": 3
}

SSE Event You Send:

event: response.output_item.done
data: {"type": "response.output_item.done", "item": {"type": "web_search_call", ...}}

What happen: Activity event is sent to GLChat

What user sees: 🔍 Searching: "AI developments 2025"

---

Step 6: Response Text (Final Output)

Your LLM generates the actual response.

Internal LLM Event:

{
    "type": "response",
    "value": "A large language model (LLM) is a neural network"
}

SSE Event You Send:

event: response.output_text.delta
data: {"type": "response.output_text.delta", "delta": "A large language model (LLM) is a neural network", ...}

What user sees: The text appears in the chat message

---

Step 7: Completion

event: response.output_text.done
data: {"type": "response.output_text.done", ...}

event: response.content_part.done
data: {"type": "response.content_part.done", ...}

event: response.output_item.done
data: {"type": "response.output_item.done", ...}

event: response.completed
data: {"type": "response.completed", "response": {...}}

What happen: Generation end

---

Event Conversion Mapping

Here's the complete mapping from internal LLM events to OpenAI format:

Thinking Events

Internal Event	OpenAI Event	Purpose
data_type: "thinking_start"	response.reasoning_summary_part.added	Initialize reasoning event
data_type: "thinking"	response.reasoning_summary_text.delta	Stream reasoning content
data_type: "thinking_end"	response.reasoning_summary_part.done	Complete reasoning display

Activity Events

Internal Event	OpenAI Event	Purpose
data_type: "activity" + type: "search"	response.output_item.done with web_search_call	Show web search
data_type: "activity" + type: "open_page"	response.output_item.done with web_search_call	Show page opened
data_type: "activity" + type: "find_in_page"	response.output_item.done with web_search_call	Show in-page search

Response Events

Internal Event	OpenAI Event	Purpose
type: "response"	response.output_text.delta	Stream text output

Example