large-language-model

Integrate Gemini 2.5 Flash multimodal LLM via gateway for text/image understanding, content generation, summarization, and multi-turn chat with SSE streaming. Use this skill whenever the user wants to

SKILL.md

GitHub

Large Language Model (Gemini 2.5 Flash)

Capability Overview

Call the Gemini 2.5 Flash multimodal large language model via the API Gateway. Supports text conversation and image understanding, with responses returned as an SSE stream.

  • Endpoint: POST https://app-dohnbj5by5mp-api-VaOwP8E7dJqa.gateway.appmedo.com/v1beta/models/gemini-2.5-flash:streamGenerateContent?alt=sse
  • Response mode: SSE streaming (Server-Sent Events)
  • Supported modalities: Multi-turn text conversation, Base64 inline images, file URI images
  • First-token latency: Up to 30 seconds; extend API timeout settings accordingly
  • Billing: This endpoint is billed per call

Request Body Structure (Overview)

Field Type Required Description
contents array Yes Array of conversation turns, each containing role and parts
contents[].role string Yes "user" or "model"
contents[].parts array Yes Message parts; each part contains one of text, inlineData, or fileData

Response Structure (SSE stream, per-frame example)

{
  "candidates": [
    {
      "content": {
        "role": "model",
        "parts": [{ "text": "That's easy! Each dog has 4 paws..." }]
      },
      "finishReason": "STOP",
      "index": 0,
      "safetyRatings": []
    }
  ]
}

Read references/stream-generate-api.md for the full parameter table, response field descriptions, and complete code for both generation-time and post-generation usage.


End-to-End Workflow

  1. The client sends the conversation history (contents array) to the Edge Function.
  2. The Edge Function reads INTEGRATIONS_API_KEY from the Deno environment, attaches the X-Gateway-Authorization header, and calls the upstream endpoint via POST.
  3. The upstream responds with an SSE stream; the Edge Function passes the stream through directly without buffering.
  4. The frontend consumes the stream with ky + eventsource-parser and renders Markdown with streamdown.

Generation-Time Usage (Agent Direct Call)

See references/stream-generate-api.md — Generation-Time Usage section.

Key points: Call the gateway directly via fetch, inject the platform key with X-Gateway-Authorization: Bearer ${apiKey}, and read the SSE stream line-by-line, parsing each frame's JSON.


Post-Generation Usage (In-App via Edge Function)

See references/stream-generate-api.md — Post-Generation Usage section.

Key points: The Edge Function passes through the upstream SSE stream; the frontend depends on ky@^1.2.3, eventsource-parser@^3.0.3, and optionally streamdown@^1.1.6 for Markdown rendering.


Notes

  • Key security: INTEGRATIONS_API_KEY may only be read server-side in the Edge Function; never expose it to the frontend.
  • Timeout: First token can take up to 30 seconds; extend the frontend request timeout accordingly (recommend ≥ 60s).
  • Error handling: Always handle 429 (quota exceeded) and 402 (insufficient balance) and display friendly messages to the user.
  • Billing: This endpoint is billed per call; avoid unnecessary repeated calls (e.g., rapid successive clicks by the user).
  • Streaming dependency: The frontend must use EventSource or eventsource-parser to consume SSE; do not use a plain fetch().json().
  • Markdown rendering: Model responses contain Markdown formatting; it is recommended to use streamdown@^1.1.6 for real-time rendering.