Gemini 2.5 — text & embeddings

Chat completions via gemini-2.5-pro / flash / flash-lite and embeddings via text-embedding-004. A single API key covers all four models.

gemini-2.5-pro gemini-2.5-flash gemini-2.5-flash-lite text-embedding-004

text_generation parameters

Endpoint: POST /v1/generations with method="text_generation". Sync mode supported — the task completes in a single HTTP request.

Name	Type	Required	Description
prompt	string	✓	User input. Non-empty.
system_instruction	string	—	System instruction — sets the model's role and style.
response_mime_type	enum	—	text/plain \| application/json. For strict JSON output, set application/json + response_schema.
response_schema	object	—	JSON schema (OpenAPI subset). Forces the model to emit valid JSON.
tools	array	—	Function calling: array of function_declarations. See example below.
temperature	number	—	0.0–2.0 (default 1.0)
max_output_tokens	integer	—	Cap on response tokens. Up to 8192.
top_p	number	—	0.0–1.0

Basic chat request

Minimal example with a system instruction. The result lives in task.result.text.

curl -X POST https://api.aigenway.com/v1/generations \
  -H "Authorization: Bearer sk_live_..." \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "google_banana",
    "model": "gemini-2.5-flash",
    "method": "text_generation",
    "params": {
      "system_instruction": "Ты — лаконичный помощник. Отвечай в одно предложение.",
      "prompt": "Почему небо голубое?",
      "temperature": 0.4,
      "max_output_tokens": 256
    }
  }'

Response shape:

{
  "task_id": "task_01HXYZ...",
  "status": "SUCCEEDED",
  "result": {
    "text": "Голубой цвет неба обусловлен рассеянием Рэлея...",
    "finish_reason": "STOP",
    "model": "gemini-2.5-flash",
    "usage": { "input_tokens": 22, "output_tokens": 31, "total_tokens": 53 },
    "files": []
  }
}

Structured JSON output

response_mime_type=application/json + response_schema guarantees task.result.text is a string you can JSON.parse directly.

curl -X POST https://api.aigenway.com/v1/generations \
  -H "Authorization: Bearer sk_live_..." \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "google_banana",
    "model": "gemini-2.5-pro",
    "method": "text_generation",
    "params": {
      "prompt": "Извлеки сущности из текста: «Apple купила за $2 млрд в 2024 году компанию Pixelmator»",
      "response_mime_type": "application/json",
      "response_schema": {
        "type": "object",
        "properties": {
          "buyer":  { "type": "string" },
          "target": { "type": "string" },
          "amount_usd": { "type": "number" },
          "year": { "type": "integer" }
        },
        "required": ["buyer", "target"]
      }
    }
  }'

Without response_schema the model may return JSON, but field shape is not guaranteed. With response_schema, fields and types are strict.

Function calling (tools)

Pass function_declarations — the model decides whether to call a function and returns arguments in tool_calls. Your code performs the call, then sends a follow-up with the result.

curl -X POST https://api.aigenway.com/v1/generations \
  -H "Authorization: Bearer sk_live_..." \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "google_banana",
    "model": "gemini-2.5-pro",
    "method": "text_generation",
    "params": {
      "prompt": "Какая сейчас погода в Москве?",
      "tools": [{
        "function_declarations": [{
          "name": "get_weather",
          "description": "Получить текущую погоду в городе",
          "parameters": {
            "type": "object",
            "properties": {
              "city": { "type": "string" }
            },
            "required": ["city"]
          }
        }]
      }]
    }
  }'

When the model decides to call a function — text is empty and tool_calls carries the name and args:

{
  "result": {
    "text": "",
    "finish_reason": "TOOL_CALL",
    "tool_calls": [
      { "name": "get_weather", "args": { "city": "Москва" } }
    ],
    "usage": { "input_tokens": 47, "output_tokens": 12, "total_tokens": 59 },
    "files": []
  }
}

Embeddings (text-embedding-004)

Endpoint: POST /v1/generations with method="embedding". Up to 100 inputs per call, native dimension 768. PER_REQUEST pricing (one inputs[] = one call).

Name	Type	Required	Description
inputs	string[]	✓	Array of strings (1–100). Each up to 2048 tokens.
task_type	enum	—	Hints the model about the downstream task. Improves retrieval quality.
output_dimensionality	integer	—	8–768. Truncate embeddings to a lower dimension (Matryoshka).
title	string	—	Document title — improves quality for RETRIEVAL_DOCUMENT.

task_type ∈ { RETRIEVAL_DOCUMENT, RETRIEVAL_QUERY, SEMANTIC_SIMILARITY, CLASSIFICATION, CLUSTERING, QUESTION_ANSWERING, FACT_VERIFICATION, CODE_RETRIEVAL_QUERY }

curl -X POST https://api.aigenway.com/v1/generations \
  -H "Authorization: Bearer sk_live_..." \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "google_banana",
    "model": "text-embedding-004",
    "method": "embedding",
    "params": {
      "inputs": [
        "Беспроводные наушники с активным шумоподавлением",
        "Шапка зимняя шерстяная мужская",
        "Кофемашина автоматическая с капучинатором"
      ],
      "task_type": "RETRIEVAL_DOCUMENT"
    }
  }'

Response shape:

{
  "task_id": "task_01HXYZ...",
  "status": "SUCCEEDED",
  "result": {
    "model": "text-embedding-004",
    "count": 3,
    "dimension": 768,
    "embeddings": [
      [0.0123, -0.0456, 0.0789, ...],
      [-0.0034, 0.0212, -0.0561, ...],
      [0.0418, -0.0102, 0.0337, ...]
    ],
    "files": []
  }
}

Truncated embeddings (Matryoshka)

output_dimensionality returns a smaller vector (8–768) — saves memory in your vector DB without losing all signal.

curl -X POST https://api.aigenway.com/v1/generations \
  -H "Authorization: Bearer sk_live_..." \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "google_banana",
    "model": "text-embedding-004",
    "method": "embedding",
    "params": {
      "inputs": ["search query: купить ноутбук asus"],
      "task_type": "RETRIEVAL_QUERY",
      "output_dimensionality": 256
    }
  }'