MelisAIEngineGemini

Google Gemini provider for the Melis AI engine — a backend engine with no back-office of its own. Package melisplatform/melis-ai-engine-gemini.

Purpose

MelisAIEngineGemini is the concrete Google Gemini implementation of MelisAIEngine's MelisAIEngineModelService contract. It supplies Gemini-specific HTTP client setup, the generateContent payload shape (the contents / parts schema), tool/function-call mapping, file handling via the File API or inline embed, and token-usage parsing. The module is stateless (no tables of its own) and ships no UI — all AI screens are provided by melis-ai / melis-ai-engine.

No React back-office

This module has no React tool and is not meant to have one — it is a backend provider layer, not a screen. It ships no ui-react/ brick, no config/react-api.php, no config/react.capabilities.php, no controller, and shows no sidebar entry or menu node in /melis-react. There are no screenshots.

In the React back-office Gemini surfaces only indirectly, as a selectable engine/model inside the MelisAI admin (Platform AI / Instances / Models & keys). When an admin creates a model whose company is Google and this module is installed, MelisAIEngine routes that model's chat and agents through the Gemini engine implemented here — with no dedicated Gemini page.

Enable it

Add to config/melis.module.load.php:

return [
    'MelisAIEngineGemini',
];

Requires melisplatform/melis-ai-engine ^6.0. The engine selects this provider automatically at runtime when the active model's company name contains Google and this module is installed; it builds MelisAIEngineModelGeminiService via $serviceManager->build(MelisAIEngineModelGeminiService::class, ['modelId' => …, 'agentId' => …]).

Key services

Service alias	Role
MelisAIEngineModelGeminiService	The Gemini provider — subclass of MelisAIEngineModelService, built per request with modelId + agentId (via MelisAIEngineModelGeminiServiceFactory).

Notable methods of MelisAIEngineModelGeminiService:

Method	Gemini implementation
setClient()	Builds a Laminas\Http\Client POSTing to {api_url}/models/{mam_generative_model}:generateContent. The API key is sent in the x-goog-api-key request header (kept out of the URL). Long timeout.
getMessageKey()	Returns 'contents'.
addToolsToPayload()	Fetches tools via getAgentFunctions(), sanitises each, renames input_schema → parameters, and wraps as tools:[{functionDeclarations:[…]}] with toolConfig.functionCallingConfig.mode='AUTO'. Falls back to a url_context tool when no functions are present.
constructContent() / addContentToPayload()	Maps role assistant → model; builds parts arrays containing {text}, {fileData:{fileUri,mimeType}} (File API), or {inlineData:{mimeType,data}} (base64).
sendCustomAI()	POSTs JSON payload (retries on 5xx / 429); parses candidates[].content.parts[] for text and functionCall entries; dispatches each tool call (MCP or built-in); appends results as {functionResponse:{name,response:{result}}} and loops until no more tool calls, a finishReason, or a safety cap.
processFiles() / processContextFiles()	Handles file uploads per mam_file_upload_mode (see File upload modes); also stores a copy via uploadDocument() for session display.
setPromptTokenCount() / setResponseTokenCount() / setTotalTokenCount()	Read usageMetadata.promptTokenCount, candidatesTokenCount, totalTokenCount.
getAllowedMimetypes()	Returns the allowed MIME types from the Gemini config block.

File upload modes

The model's mam_file_upload_mode controls how files reach Gemini:

• fileapi (default) — uploads to the Gemini File API via a multipart POST, then polls until the file state is ACTIVE, and references the returned fileApiUri.
• embed — sends file bytes inline as base64 (inlineData), with text extraction for supported formats.

Configuration

Fixed provider config lives under config['plugins']['melisaiengine']['datas']['AI']['Gemini'] (in config/app.interface.php, no UI):

Key	Purpose
api_url	Base URL for generateContent requests (https://generativelanguage.googleapis.com/v1beta).
upload_url	Gemini File API upload endpoint (https://generativelanguage.googleapis.com/upload/v1beta/files).
allowed_mimetypes	Accepted MIME types (returned by getAllowedMimetypes()).

Per-model values come from the model row configured in MelisAI, not from this module:

• mam_generative_model — the Gemini model id (e.g. gemini-…), used in the request URL.
• mapk_keys — the Google API key, sent in the x-goog-api-key header. It comes from the model's platform key record, is not stored in config files, and has no environment-variable fallback.
• mam_file_upload_mode (fileapi default, or embed) and mam_internal_upload — govern how files are uploaded.

API key in header

Unlike earlier releases that appended the key as a ?key=… query parameter, the current source sends the Google API key in the x-goog-api-key request header, keeping it out of URLs and proxy logs.

Tool / function-call mapping

Gemini uses parameters where the engine contract uses input_schema; addToolsToPayload() performs this rename before sending. Tool calls arrive as {functionCall:{name,args}} with no id (positional, unlike Claude's tool_use_id). The service routes by name via MelisAIEngineMcpService::isMcpTool() and returns {functionResponse:{name,response:{result:…}}}.

Differences vs Claude provider

Aspect	Gemini	Claude
Message key	contents	messages
Assistant role label	model	assistant
Tool schema key	parameters	input_schema
Tool-call id	none (positional)	tool_use_id
Auth method	x-goog-api-key header	x-api-key header
Default file mode	fileapi	embed

Key files

Concern	Path
Provider service	vendor/melisplatform/melis-ai-engine-gemini/src/Service/MelisAIEngineModelGeminiService.php
Service factory	vendor/melisplatform/melis-ai-engine-gemini/src/Service/Factory/MelisAIEngineModelGeminiServiceFactory.php
Fixed config	vendor/melisplatform/melis-ai-engine-gemini/config/app.interface.php

See also

• MelisAIEngine — the abstract contract and shared runtime.
• MelisAIEngineClaude — the Anthropic provider.
• MelisAI — back-office for agents, instances and models.