Gemini Batch API Skill

Large-scale asynchronous document processing using Google's Gemini models.

When to Use

• Process thousands of documents with the same prompt
• Cost-effective bulk extraction (50% cheaper than synchronous API)
• Jobs that can tolerate 24-hour completion windows

IRON LAW: Use Examples First, Never Guess API

YOU MUST READ EXAMPLES BEFORE WRITING ANY CODE. NO EXCEPTIONS.

The Rule

User asks for batch API work
    ↓
MANDATORY: Read examples/batch_processor.py or examples/icon_batch_vision.py
    ↓
Copy the pattern exactly
    ↓
DO NOT guess parameter names
DO NOT try wrapper types
DO NOT improvise API calls

Why This Matters

The Batch API has non-obvious requirements that will fail silently:

1. Metadata must be flat primitives - Nested objects cause cryptic errors
2. Parameter is dest= not destination= - Wrong name → TypeError
3. Config is plain dict - Not a wrapper type
4. Examples are authoritative - Working code beats assumptions

Rationale: Previous agents wasted hours debugging API errors that the examples would have prevented. The patterns in examples/ are battle-tested production code.

Rationalization Table - STOP If You Think:

Excuse	Reality	Do Instead
"I know how APIs work"	This API has non-obvious gotchas	Read examples first
"I can figure it out"	You'll waste 30+ minutes on trial-and-error	Copy working patterns
"The examples might be outdated"	They're maintained and tested	Trust the examples
"I need to customize anyway"	Customization comes AFTER copying base pattern	Start with examples, then adapt
"Reading examples takes too long"	2 minutes to read saves 30 minutes debugging	Read examples first
"My approach is simpler"	Simpler approaches already failed	Use proven patterns

Red Flags - STOP If You Catch Yourself:

• "Let me try destination= instead of dest=" → WRONG. Read examples.
• "I'll create a CreateBatchJobConfig object" → WRONG. Plain dict only.
• "I'll nest metadata like a normal API" → WRONG. Flat primitives only.
• "This should work like other Google APIs" → WRONG. Batch API is different.
• "I'll figure out the JSONL format" → WRONG. Copy from examples.

MANDATORY Checklist Before ANY Batch API Code

• Read examples/batch_processor.py OR examples/icon_batch_vision.py
• Identify which example matches your use case (Standard API vs Vertex AI)
• Copy the example's API call pattern exactly
• Copy the example's JSONL structure exactly
• Copy the example's metadata structure exactly
• Only THEN adapt for your specific needs

Enforcement: If you write batch API code without reading examples first, you are violating this IRON LAW and WILL encounter preventable errors.

Prerequisites

Install gcloud SDK

# macOS
brew install google-cloud-sdk

# Linux
curl https://sdk.cloud.google.com | bash

Authentication Setup

# Step 1: User authentication (for gcloud commands)
gcloud auth login

# Step 2: Application Default Credentials (for Python libraries)
gcloud auth application-default login

# Step 3: Enable Vertex AI API
gcloud services enable aiplatform.googleapis.com

Why both auth methods?

• gcloud auth login: For gsutil and gcloud CLI commands
• gcloud auth application-default login: For google-generativeai Python library
• CRITICAL: Vertex AI requires ADC (step 2), not just API key

Create GCS Bucket

# Create bucket in us-central1 (required region)
gsutil mb -l us-central1 gs://your-batch-bucket

# Verify location
gsutil ls -L -b gs://your-batch-bucket | grep "Location"

See references/gcs-setup.md for complete setup guide.

Quick Start

Standard Gemini API (API Key)

from examples.batch_processor import GeminiBatchProcessor

processor = GeminiBatchProcessor(
    bucket_name="my-batch-bucket",  # Must be in us-central1
    model="gemini-2.0-flash-lite"
)

results = processor.run_pipeline(
    input_dir="./documents",
    prompt="Extract as JSON: {title, date, summary}",
    output_dir="./results"
)

Vertex AI (Recommended)

import google.generativeai as genai

# Use Vertex AI with ADC
client = genai.Client(
    vertexai=True,
    project="your-project-id",
    location="us-central1"
)

# Submit batch job
job = client.batches.create(
    model="gemini-2.5-flash-lite",
    src="gs://bucket/requests.jsonl",
    dest="gs://bucket/outputs/"
)

Core Workflow

1. Upload files to GCS bucket (us-central1 region required)
2. Create JSONL request file with document URIs and prompts
3. Submit batch job via genai.batches.create()
4. Poll for completion (jobs expire after 24 hours)
5. Download and parse results from output URI
6. Handle failures gracefully (partial failures are common)

IRON LAW: Metadata and API Call Structure

METADATA MUST BE FLAT PRIMITIVES. API PARAMETERS ARE SIMPLE STRINGS.

Rule 1: Metadata Structure

CORRECT ✓
"metadata": {
    "request_id": "icon_123",        # String
    "file_name": "copy.svg",         # String
    "file_size": 1024                # Integer
}

WRONG ✗
"metadata": {
    "request_id": "icon_123",
    "file_info": {                   # ← NESTED OBJECT FAILS!
        "name": "copy.svg",
        "size": 1024
    }
}

WORKAROUND (if complex data needed)
"metadata": {
    "request_id": "icon_123",
    "file_info": json.dumps({"name": "copy.svg", "size": 1024})  # JSON string OK
}

Why: Vertex AI stores metadata in BigQuery-compatible format. BigQuery doesn't support nested types. Violation causes: "metadata" in the specified input data is of unsupported type.

Rule 2: API Call Structure

CORRECT ✓
job = client.batches.create(
    model="gemini-2.5-flash-lite",
    src="gs://bucket/input.jsonl",        # Just a string
    dest="gs://bucket/output/",           # Just a string
    config={"display_name": "my-job"}     # Just a dict
)

WRONG ✗
job = client.batches.create(
    model="gemini-2.5-flash-lite",
    src="gs://bucket/input.jsonl",
    destination="gs://bucket/output/",    # ← PARAMETER DOESN'T EXIST!
)

WRONG ✗
job = client.batches.create(
    model="gemini-2.5-flash-lite",
    src="gs://bucket/input.jsonl",
    config=types.CreateBatchJobConfig(    # ← DON'T INSTANTIATE TYPES!
        dest="gs://bucket/output/"
    )
)

Why: The SDK uses simple types. Parameter is dest= (not destination). Config is a plain dict (not a type instance). The SDK converts internally.

Rationalization Table - STOP If You Think:

Excuse	Reality	Do Instead
"Nested metadata is cleaner"	It fails silently with cryptic errors	Flatten or use json.dumps()
"I'll try destination= parameter"	Parameter doesn't exist, causes TypeError	Use dest=
"I should use CreateBatchJobConfig"	That's internal typing, not for you	Pass plain dict to config=
"Other APIs accept nested objects"	This one doesn't, it's BigQuery-backed	Follow the examples
"I'll fix it if it breaks"	Job fails 5 minutes after submission	Get it right the first time

Pre-Submission Validation

# Add this check BEFORE submitting batch job
def validate_metadata(metadata: dict):
    """Ensure metadata contains only primitive types."""
    for key, value in metadata.items():
        if isinstance(value, (dict, list)):
            raise ValueError(
                f"Metadata '{key}' is {type(value).__name__}. "
                f"Only primitives (str, int, float, bool) allowed. "
                f"Use json.dumps() for complex data."
            )
        if not isinstance(value, (str, int, float, bool, type(None))):
            raise ValueError(f"Unsupported type for '{key}': {type(value)}")

# Use it:
for request in batch_requests:
    validate_metadata(request["metadata"])

Enforcement: Jobs WILL fail if metadata contains nested objects. There is no way around this requirement.

Key Gotchas

Issue	Solution
Nested metadata fails	Use flat primitives or json.dumps() for complex data
TypeError: unexpected keyword	Use dest= not destination=, pass plain dict
Auth errors with Vertex AI	Run gcloud auth application-default login
vertexai=True requires ADC	API key is ignored with vertexai=True
Missing aiplatform API	Run gcloud services enable aiplatform.googleapis.com
Region mismatch	Use us-central1 bucket only
Wrong URI format	Use gs:// not https://
Invalid JSONL	Use scripts/validate_jsonl.py
Image batch: inline data	Use fileData.fileUri for batch, not inline
Duplicate IDs	Hash file content + prompt for unique IDs
Large PDFs fail	Split at 50 pages / 50MB max
JSON parsing fails	Use robust extraction (see gotchas.md)
Output not found	Output URI is prefix, not file path

Top 2 mistakes (bolded above):

1. Using nested objects in metadata instead of flat primitives
2. Guessing parameter names instead of using dest=

See references/gotchas.md for detailed solutions (now with Gotchas 10 & 11).

Rate Limits

Limit	Value
Max requests per JSONL	10,000
Max concurrent jobs	10
Max job size	100MB
Job expiration	24 hours

Recommended Models

Model	Use Case	Cost
gemini-2.0-flash-lite	Most batch jobs	Lowest
gemini-2.0-flash	Complex extraction	Medium
gemini-1.5-pro	Highest accuracy	Highest

Additional Resources

References

• references/gcs-setup.md - NEW: Complete GCS and Vertex AI setup guide
• references/gotchas.md - 9 critical production gotchas (updated auth section)
• references/best-practices.md - Idempotent IDs, state tracking, validation
• references/troubleshooting.md - Common errors and debugging
• references/vertex-ai.md - Enterprise alternative with comparison
• references/cli-reference.md - gsutil and gcloud commands

Examples

• examples/icon_batch_vision.py - NEW: Batch vision analysis with Vertex AI
• examples/batch_processor.py - Complete GeminiBatchProcessor class
• examples/pipeline_template.py - Customizable pipeline template

Scripts

• scripts/validate_jsonl.py - Validate JSONL before submission
• scripts/test_single.py - Test single request before batch

External Documentation

• Gemini Batch API Guide
• Google Cloud Storage
• Vertex AI Batch Prediction

Date Awareness

Pattern from oh-my-opencode: Gemini API and documentation evolve rapidly.

Current date: Use datetime.now() for:

• API version checking
• Model availability ("gemini-2.5-flash-lite available as of Dec 2024")
• Documentation freshness validation

When in doubt about API features or model names, verify against current date and check latest Gemini API docs.