OpenRouter AI Plugin Reference Manual

1. Overview

Plugin Name: OpenRouter AI

Type: Cloud AI / Multimodal

Identifier: uniflow.plugin.openrouter

Platform: CrossPlatform (Windows & Linux x64/ARM64)

Description

Connects Uniflow directly to cloud-hosted Multimodal Artificial Intelligence models (Vision, Audio, Text, PDF, and Video) via the unified OpenRouter REST API gateway. The plugin allows industrial rule workflows to execute visual inspection, PPE safety monitoring, license plate object detection, voice command recognition, audio analysis, PDF document data extraction, and AI media generation (Image/Video/TTS) using state-of-the-art AI models (e.g. OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, Google Gemini 1.5 Pro/Flash, Meta Llama 3.2 Vision, Luma Photon, Whisper).

2. Technical Architecture

The OpenRouter AI plugin provides a 100% self-contained integration layer between the Uniflow Rule Engine and cloud inference providers.

┌─────────────────────────────────────────────────────────────────────────────────┐
│                              UNIFLOW RULE ENGINE                                │
└────────────────────────────────────────┬────────────────────────────────────────┘
                                         │
                                         ▼
┌─────────────────────────────────────────────────────────────────────────────────┐
│                            Uniflow.Plugin.OpenRouter                            │
│  ┌───────────────────────┐ ┌───────────────────────┐ ┌───────────────────────┐  │
│  │ OpenRouterMediaResolver│ │   OpenRouterApiClient │ │ Base64LogSanitizer    │  │
│  │ (ImageSharp Scaler)   │ │ (Retry + Fallback)    │ │ (Log Flooding Shield) │  │
│  └───────────────────────┘ └───────────────────────┘ └───────────────────────┘  │
│  ┌───────────────────────────────────────────────────────────────────────────┐  │
│  │ Supervisor Loop (Periodic 30s Health, Credit Usage & Rate Limits)          │  │
│  └───────────────────────────────────────────────────────────────────────────┘  │
└────────────────────────────────────────┬────────────────────────────────────────┘
                                         │
                                         ▼
┌─────────────────────────────────────────────────────────────────────────────────┐
│                    OpenRouter Gateway (https://openrouter.ai/api/v1)            │
└────────────────────────────────────────┬────────────────────────────────────────┘
                                         │
       ┌───────────────────┬─────────────┴──────┬───────────────────┐
       ▼                   ▼                    ▼                   ▼
┌──────────────┐    ┌──────────────┐     ┌──────────────┐    ┌──────────────┐
│ OpenAI GPT-4o│    │Claude 3.5    │     │ Gemini 1.5   │    │  Whisper /   │
│ Vision & Audio    │ Sonnet Vision│     │  Pro & Flash │    │ Speech-to-Text│
└──────────────┘    └──────────────┘     └──────────────┘    └──────────────┘

Core Architectural Features

2.1 Memory Safety, Resource Efficiency & Log Protection

To ensure high-throughput execution in 24/7 industrial production environments, the plugin implements strict memory allocation boundaries:

1. Chunked Streaming File Downloads: FileDownloadQueueManager downloads files directly from the network stream to disk (FileStream) using fixed 8 KB memory buffers via HttpCompletionOption.ResponseHeadersRead. Memory consumption remains O(1) constant (~8 KB per download) regardless of whether the file size is 10 MB or 5 GB.

2. 50 MB Inline Base64 File Size Ceiling: When reading local non-image media files (audio, video, PDF) into memory for inline API transport, OpenRouterMediaResolver enforces a strict 50 MB safety ceiling to prevent Large Object Heap (LOH) fragmentation and memory spikes.

3. ImageSharp Image Resizing & Allocation Controls: Camera snapshot images are downscaled to MaxImageDimension in memory and disposed immediately using using var image = await Image.LoadAsync(...). Memory streams and image memory buffers are reclaimed by GC promptly. MaxImageDimension is fully configurable in the Source Editor UI (default 1024 px, editable up to 4096 px, or set to 0 to disable downscaling and preserve 100% original full resolution).

4. Log Shielding (Base64LogSanitizer): To avoid polluting disk storage and CPU overhead during DBG log generation, regex sanitizers replace raw Base64 data strings with lightweight string placeholders prior to writing log entries.

System Interaction & Exposed Catalog Routes

The OpenRouter AI plugin interfaces with OpenRouter's LLM gateway (https://openrouter.ai/api/v1/chat/completions). It builds multimodal payloads (text prompts, Base64 images, audio, PDF files), enforces strict JSON Schema outputs, and returns structured data.

Catalog Routes & Node Integration

Input Event Triggers (Input Nodes):

Executable Actions (Action Nodes):

Architecture Diagram

3. Configuration Parameters

The following configuration settings are available in the WPF User Interface (OpenRouterSourceEditorView) when creating or configuring an instance of this plugin source:

3.1 Preset Prompt Templates & Dynamic UI Schema Visibility

The plugin allows administrators to build a table of Preset Prompt Templates stored inside the plugin configuration. Each preset is assigned a PresetType:

3.2 Model Endpoints Catalog Helper Window & Dynamic Parameter Binding

Because supported resolutions, aspect ratios, streaming capabilities, and pricing vary per provider endpoint, the plugin features a dedicated 3-Column Model Endpoints Catalog Helper Window (OpenRouterModelCatalogDialog):

1. Access: Click the Helper button (🔍) on the OpenRouter source row in the Sources list, "🌐 Browse Model Endpoints Catalog..." in the Source Editor view, or "🌐 Inspect Endpoints" in the Preset Edit dialog.

2. Layout & Capabilities:

     {
       "id": "bytedance-seed/seedream-4.5",
       "endpoints": [
         {
           "provider_name": "Bytedance",
           "provider_slug": "bytedance",
           "provider_tag": "bytedance",
           "supported_parameters": {
             "resolution": { "type": "enum", "values": ["1K", "2K", "4K"] },
             "seed": { "type": "boolean" }
           },
           "allowed_passthrough_parameters": [],
           "supports_streaming": false,
           "pricing": [
             { "billable": "output_image", "unit": "image", "cost_usd": 0.05 }
           ]
         }
       ]
     }
     

3. Background Auto-Refresh (30 Minutes): Endpoint metadata is automatically fetched from OpenRouter GET /api/v1/models every 30 minutes in the background, or refreshed immediately on demand via Refresh Endpoint Catalog.

4. Dynamic Output Target Parameter Binding: When configuring presets or Output Target nodes, the plugin queries the cached endpoint specs for the target model and populates resolution (["1K", "2K", "4K"] or ["1024x1024", "512x512"]), aspect_ratio, and seed options dynamically.

4. Exposed Routes & Data Types

This plugin exposes catalog fields across Input Source (telemetry data), Output Target (action sinks), and Event Input (trigger sources) rule graph nodes:

4.1 Input Source Telemetry Fields

The Input Source node reads live operational metrics and credit usage from the OpenRouter gateway:

4.2 Output Target Action Nodes

The Output Target node acts as an action sink to execute AI multimodal requests or enqueue file downloads:

4.3 Event Input Trigger Nodes

The Event Input node triggers downstream rule logic upon completion of an AI preset prompt operation or background file download:

4.4 Event Output Telemetry Fields

When an Event Input node triggers, it exposes the following output fields for rule execution:

Preset Completion Events (event:preset_<id>)

File Downloaded Event (event:file_downloaded)

5. Usage Examples & Workflows

5.1 AI Image Generation: 2-Step Request ➔ Event & Download to Disk Workflow

In asynchronous AI image generation workflows, execution happens in two distinct steps across two rules:

1. Rule 1 (Send Request): An industrial trigger (e.g., Modbus coil, camera sensor, or timer) dispatches the image generation request with specific parameters (prompt, resolution, aspect_ratio, style, payload).

2. Rule 2 (Receive Result & Save to Disk): When the OpenRouter API finishes generating the image, it raises a completion event (event:preset_<id> or event:generate_image). Rule 2 catches this event, parses StructuredResultJson, extracts the image URL or Base64 Data URI, and triggers OpenR / Download File to write the image file directly to disk (D:\_AiGen\ImageGen.jpg).

┌─────────────────────────────────────────────────────────────────────────────────────────┐
│                                RULE 1: Send Request                                     │
│  [Input Source: LocalModbus] ──(Trigger=true)──► [Output Target: OpenR / ImageGen]      │
│  COIL_0                                          - prompt: "Generate a futuristic image..." │
│                                                  - resolution: "256x256"                │
│                                                  - aspect_ratio: "1:1"                  │
│                                                  - style: "futuristic"                  │
│                                                  - payload: "ASDF1234"                  │
└─────────────────────────────────────────────────────────────────────────────────────────┘
                                           │
                                           ▼ (Asynchronous OpenRouter API Inference)
┌─────────────────────────────────────────────────────────────────────────────────────────┐
│                           RULE 2: Receive Event & Download File                         │
│  [Event Input: OpenR] ─────► [JSON Deserialize: OpenRouter] ──► [Output Target: OpenR]  │
│  event:preset_<id>           - Reads: StructuredResultJson      OpenR / Download File   │
│  Outputs:                    - Success ──(Trigger=true)───────► - url: connected from   │
│  - Success                   - Url ────(File URL)─────────────►   JSON Deserialize.Url  │
│  - Payload ("ASDF1234")                                         - path: "D:\_AiGen"     │
│  - StructuredResultJson                                         - filename: "ImageGen.jpg"│
│                                                                 - payload: "Download_1" │
└─────────────────────────────────────────────────────────────────────────────────────────┘
                                           │
                                           ▼
                              Saved to Disk: D:\_AiGen\ImageGen.jpg

Step 1: Rule 1 Graph Configuration ("0_CoilTriggerGenImageAI")

Step 2: Rule 2 Graph Configuration ("1_ImageGenerated")

1. When OpenRouter completes generation, it publishes event:preset_<id>.

2. JSON Deserialize reads StructuredResultJson, outputting Success = true and Url = "data:image/jpeg;base64,...".

3. Output Target: Download File receives Trigger = true, decodes the Base64 Data URI, and saves the image directly to D:\_AiGen\ImageGen.jpg.

4. The plugin publishes event:file_downloaded with DownloadStatus = Success and FullFilePath = D:\_AiGen\ImageGen.jpg.

5.2 Custom JSON Schema Enforcement

In physical security, perimeter surveillance, and industrial access control environments, AI models can analyze audio recordings, camera snapshots, or video clips and respond with structured JSON output enforcing a specific schema.

Step 1: Configure Schema Format Mode in Preset Editor

When creating a preset (e.g., PresetType.AudioUnderstanding, PresetType.ImageUnderstanding, or PresetType.VideoUnderstanding):

1. In the Preset Editor, set Format Mode to JsonObjectRef (or BuiltInJsonObject).

2. Select your custom JSON Object schema created in Uniflow's "Json Objects" manager.

3. The plugin automatically attaches response_format: { "type": "json_object" } to the OpenRouter request and appends JSON formatting instructions.

Security Scenario A: Audio Security Threat Analysis (Perimeter Microphone / Intercom)

An audio microphone near a restricted fence line records an audio clip (D:\_AiGen\tts.mp3). The AI determines if a security threat exists.

Custom JSON Object Schema (Security Audio Threat Analysis):

{
  "ThreatDetected": true,
  "ThreatCategory": "GlassBreakage",
  "Confidence": 0.95,
  "SoundEvents": "Shattering glass followed by running footsteps",
  "UrgencyLevel": "High",
  "RecommendedAction": "Dispatch Security Patrol to Zone 1"
}

Preset Prompt Template (Configured in Preset Editor for Perimeter_Audio_Guard):

Analyze the attached audio recording from the Zone 1 Perimeter Microphone.
Identify all sound events, assess if there is a security threat or emergency, and classify the threat type (e.g., GlassBreakage, FenceClimbing, VocalAggression, Gunshot, FalseAlarm).

Respond ONLY in valid JSON matching the schema format:
{
  "ThreatDetected": boolean,
  "ThreatCategory": string,
  "Confidence": number,
  "SoundEvents": string,
  "UrgencyLevel": "Low" | "Medium" | "High" | "Critical",
  "RecommendedAction": string
}

User Input Parameters (Sent in Output Target Action Node):

Rule Execution Flow:

1. Rule 1 (Send Audio): Triggered by perimeter sound sensor → Dispatches Output Target: OpenR / Perimeter_Audio_Guard with the input parameters above.

2. Rule 2 (Evaluate Threat & Fire Alarm): Listens for Event Input: OpenR / Perimeter_Audio_Guard Completed → Connects StructuredResultJson to JSON Deserialize (Security Audio Threat Analysis).

3. Downstream logic: If ThreatDetected == true AND UrgencyLevel == "High" → Triggers TCP / Send Data to transmit an alarm to the central VMS control room.

Security Scenario B: Gate PPE & Vehicle Inspection (Access Control Camera)

A fixed camera at a security entrance gate captures a worker and vehicle (D:\Snapshots\gate_worker.jpg).

Custom JSON Object Schema (Gate PPE Safety Inspection):

{
  "AccessGranted": false,
  "HardHatPresent": true,
  "SafetyVestPresent": false,
  "PersonCount": 1,
  "ViolationDetails": "Worker is missing a high-visibility safety vest",
  "VehiclePlate": "B-102-XYZ"
}

Preset Prompt Template (Configured in Preset Editor for Gate_PPE_Inspection):

Inspect the attached security camera snapshot of the entrance gate.
Check if all visible personnel are wearing required Personal Protective Equipment (Hard Hat and High-Vis Safety Vest). Also extract any visible vehicle license plate.

Respond in valid JSON:
{
  "AccessGranted": boolean,
  "HardHatPresent": boolean,
  "SafetyVestPresent": boolean,
  "PersonCount": integer,
  "ViolationDetails": string,
  "VehiclePlate": string
}

User Input Parameters (Sent in Output Target Action Node):

Rule Execution Flow:

1. Rule 1: Gate motion sensor triggers snapshot → Dispatches Output Target: OpenR / Gate_PPE_Inspection with the input parameters above.

2. Rule 2: Listens for Gate_PPE_Inspection Completed → JSON Deserialize (Gate PPE Safety Inspection).

3. If AccessGranted == true → Set LocalModbus / GATE_BARRIER coil to true (opens gate).

4. If AccessGranted == false → Sound strobe warning light and log ViolationDetails.

Security Scenario C: Video CCTV Intrusion Analysis (Surveillance Camera Clip)

A VMS plugin captures a 5-second video recording clip (D:\VideoClips\camera_4_intrusion.mp4) when a motion perimeter is breached after hours.

Custom JSON Object Schema (CCTV Intrusion Event Analysis):

{
  "IntrusionConfirmed": true,
  "SubjectType": "Human",
  "SubjectCount": 2,
  "Behavior": "Scaling perimeter fence and carrying a backpack",
  "BoundingBoxLocation": "Upper Right Fence Quadrant",
  "AlarmSeverity": 5
}

Preset Prompt Template (Configured in Preset Editor for CCTV_Fence_Guard):

Analyze the attached surveillance video recording clip captured from Camera 4 at 02:00 AM.
Determine if there is an active perimeter intrusion by humans or vehicles. Ignore animals (cats, dogs, birds) or wind-blown objects.

Respond in valid JSON:
{
  "IntrusionConfirmed": boolean,
  "SubjectType": "Human" | "Vehicle" | "Animal" | "None",
  "SubjectCount": integer,
  "Behavior": string,
  "BoundingBoxLocation": string,
  "AlarmSeverity": integer
}

User Input Parameters (Sent in Output Target Action Node):

Scenario A: Image Object Detection & Safety Inspection

To detect whether safety helmets and reflective vests are present in an image captured from an ONVIF camera:

1. Configure an ONVIF plugin rule to save snapshot images to C:\Snapshots\gate1.jpg.

2. Add an Output Target Action Node targeting openrouter:understand_image with:

3. Add an Event Input Node listening to event:understand_image.

4. Connect downstream conditional nodes evaluating HelmetDetected == false to sound a local warning alarm.

Logic Flow Diagram:

Scenario B: Voice Command Recognition (Speech-to-Text)

To process audio voice recordings captured from an intercom or PA microphone:

1. Ingest the audio recording file path (e.g. C:\Recordings\command.wav).

2. Add an Output Target Action Node targeting openrouter:speech_to_text with media_input = C:\Recordings\command.wav.

3. Add an Event Input Node listening to event:speech_to_text.

4. Use the Transcript field output in downstream string matching nodes (str_contains) to trigger industrial machinery actions.

Logic Flow Diagram:

Scenario C: Dynamic Prompt Overriding from Rule Graph

To dynamically inject custom inspection instructions built by an upstream rule node:

1. Construct a dynamic prompt string (e.g. using a string concatenation node: "Inspect cargo container " + container_id + " for physical damage").

2. Connect the string output port of the concatenation node into the prompt input port of an Output Target Action Node configured for openrouter:preset_cargo_inspector.

3. The plugin will execute the call using the connected dynamic prompt, overriding the preset's default template.

Logic Flow Diagram:

Scenario D: Two-Rule AI Generation & Asynchronous Model/File Download Workflow

To generate a model/image asset via OpenRouter and automatically download the resulting remote file:

Rule 1 (Generation Request):

1. Add an Output Target Action Node configured for openrouter:generate_image (or openrouter:generate_video).

2. Provide prompt = "Industrial warning sign icon, high resolution".

3. Execution dispatches the request to OpenRouter.

Rule 2 (Event Completion & Download Dispatch):

1. Add an Event Input Node listening for event:generate_image.

2. Evaluate Success == true.

3. Extract the generated media URL from Url output port (or StructuredResultJson).

4. Add an Output Target Action Node targeting openrouter:download_file (or openrouter:download_model) with parameters:

5. The download queue manager ensures unique URL handling and executes up to 5 retry attempts logging progress at DBG level.

6. When complete, an event:file_downloaded event is published with DownloadStatus = Success and FullFilePath = C:\Uniflow\Downloads\warning_sign.png.

Logic Flow Diagram: