robotframework-gemini

English | Português (Brasil)

Robot Framework keyword library for Google Gemini oracles: text-only evaluation (API payloads, logs, JSON, etc.) or multimodal checks with an image (PNG file or a Robot Framework Browser screenshot).

• Multimodal: sends prompt text and the capture as a byte Part (google-genai).
• Recommended flow: two inputs — context (test framing) and evaluation (criterion or yes/no question).
• Browser optional: screenshot keywords need Library Browser; text keywords work without a browser.

Installation

pip install robotframework-gemini

With Robot Framework Browser (Playwright) screenshot support:

pip install "robotframework-gemini[browser]"
python -m pip install robotframework

Local development:

python -m pip install -e ".[dev]"

Environment variables and Library import

By default the Library reads key and model from the environment. You can also pass them on import (useful in CI or suites that keep credentials in Robot variables):

*** Variables ***
${GEMINI_API_KEY}    %{GEMINI_API_KEY}
${GEMINI_MODEL}      gemini-flash-latest

*** Settings ***
Library    GeminiLibrary    api_key=${GEMINI_API_KEY}    model=${GEMINI_MODEL}

Model only (key still comes from the environment):

Library    GeminiLibrary    model=gemini-flash-latest

Explicit package import (equivalent):

Library    robotframework_gemini.library.GeminiLibrary    api_key=${GEMINI_API_KEY}    model=${GEMINI_MODEL}

Do not commit a literal API key; prefer %{GEMINI_API_KEY} or CI secrets.

Keyword documentation

• English: docs/KEYWORDS.en.md
• Português: docs/KEYWORDS.pt-BR.md

Usage in Robot Framework

Note: examples use Set Variable and Catenate instead of the VAR keyword (Robot Framework 7.0+) to stay compatible with older Robot Framework versions.

Text only (no Browser)

*** Settings ***
Library    GeminiLibrary

*** Keywords ***
Validate API response
    ${context}=       Set Variable    {"status": "ok", "items": 3}
    ${evaluation}=    Set Variable    Does the payload indicate success with items?
    ${model_response}=    Gemini Evaluate Text    ${context}    ${evaluation}
    Log    ${model_response}

With a screenshot (Browser)

Import Browser before Gemini:

*** Settings ***
Library    Browser
Library    GeminiLibrary

*** Keywords ***
Check screen against a neutral criterion
    ${context}=       Set Variable    List filtered by Active status.
    ${evaluation}=    Set Variable    Do all visible items show Active status?
    ${model_response}=    Gemini Evaluate With Screen    ${context}    ${evaluation}
    Log    ${model_response}

Recommended import (RobotCode, runtime, and PyPI):

Library    GeminiLibrary    api_key=${GEMINI_API_KEY}    model=${GEMINI_MODEL}

Explicit package path (IDEs that prefer a full Python path):

Library    robotframework_gemini.library.GeminiLibrary    api_key=${GEMINI_API_KEY}    model=${GEMINI_MODEL}

IDE and RobotCode

In a repository clone, without installing:

• robot.toml — python-path = ["src"] for RobotCode/LSP
• .vscode/settings.json — extraPaths for Pylance/Cursor

With a local venv: pip install -e ".[dev]", then reload the IDE window after changes.

With an image file already saved:

Browser.Take Screenshot    ${OUTPUT_DIR}/screen.png
${model_response}=    Gemini Evaluate With Image File    ${context}    ${evaluation}    ${OUTPUT_DIR}/screen.png

Verdict via prompt (first line) and 1–5 score:

${model_response}=    Gemini Evaluate With Screen    ${context}    ${evaluation}
...    extra_instructions=Reply with one word on the first line: Yes or No.
${verdict}=    Get Line    ${model_response}    0
Should Be Equal As Strings    ${verdict}    Yes

# Score 1–5: two steps (log the reason) or one-line shortcut
${model_response}=    Gemini Evaluate Text Rating    ${context}    ${evaluation}
${rating_score}=    Gemini Parse Rating    ${model_response}

${rating_score}=    Gemini Evaluate Text Rating Score    ${context}    ${evaluation}

Details for the three rating keywords: Scores 1–5 (EN).

See also examples/demo_template.robot.

Related projects

Part of a Robot Framework tooling set:

Tests

pytest

Tests mock generate_content; there are no real API calls.

Security

See SECURITY.md for how to report vulnerabilities privately.

License

MIT.