jev-chat-jarvis

### Deep Dive: Architecture and Setup of Jev Chat Jarvis If you are looking for an Android-native chat assistant that avoids invasive Xposed hooks, modified APKs, or direct database reads, `jev-chat-jarvis` takes a clean, passive approach. Designed for Android 11+, it functions as an on-device "conversation copilot" that uses standard Android Accessibility Services to analyze on-screen chat text without touching target application binaries. --- ### System Architecture and Data Pipeline `jev-chat-jarvis` operates entirely through a read-only execution flow. Instead of hooking into private APIs or intercepting network traffic, it relies on UI node parsing and local processing. ``` ┌──────────────────┐ ┌───────────────────┐ ┌───────────────────────┐ │ Target App UI │ ──> │ App Adapter / OCR │ ──> │ Context Enrichment │ │ (QQ, X, Feishu) │ │ (Parses Screen) │ │ (Local KB & Contacts) │ └──────────────────┘ └───────────────────┘ └───────────┬───────────┘ │ ┌──────────────────┐ ┌───────────────────┐ │ │ Floating Overlay │ <── │ Reply Model │ <───────────────┘ │ (Manual Send) │ │ (3 Candidates) │ │ └──────────────────┘ └───────────────────┘ ┌───┴───────────────────┐ │ Judgment Model │ │ (Intent & Danger Level│ └───────────────────────┘ ``` #### 1. Ingestion and Platform Adapters The core runtime reads UI elements currently visible on the screen. Platform-specific adapters parse these UI elements: * QQ & X: Text elements are parsed via accessibility node trees. * Feishu: Uses local OCR to capture chat content when standard node text is inaccessible. * Extensibility: Support for new applications requires writing an app adapter (typically around a few dozen lines of code). Note that Android WeChat support has been explicitly removed and is not supported. #### 2. Context Injection Before sending payloads to LLM providers, the app queries its private on-device storage. If a contact profile or relevant entry from your local knowledge base matches the current context, that background information is automatically appended to the context payload. #### 3. Two-Stage AI Analysis Unlike simple single-prompt wrappers, the app decouples decision-making into distinct steps: 1. Judgment Model: Analyzes the counterparty's real intent, evaluates danger/risk levels, and determines if a response is immediately required. 2. Reply Model: Takes the intent/risk analysis and drafts three prioritized candidate replies. #### 4. UI Overlay and Input Auto-Fill The generated candidate replies appear on a floating overlay window above your chat application. Tapping a candidate uses accessibility APIs to insert the text directly into the target app's input field. The application never triggers the send action automatically, nor does it interact with sensitive UI components like money transfers or red envelopes. --- ### Setup and Configuration Because there is no backend middleman server hosted by the developer, requests are dispatched directly from your device to your chosen model providers. #### Configurable API Endpoints The client requires you to configure three distinct API slots in the settings panel: * Judgment API Slot: Endpoint and key for intent analysis and danger evaluation. * Reply API Slot: Endpoint and key for candidate text generation. * Vision API Slot: Endpoint for OCR tasks (used for apps like Feishu). #### Installation Steps 1. Download `jev-assistant-v1.4-release.apk` (requires Android 11 or higher). 2. Install the APK on your device and grant the necessary Android Accessibility Service permissions so the app can read screen text and auto-fill the input target. 3. Open the app settings and input your respective API credentials for the Judgment, Reply, and Vision slots. 4. Populate your local knowledge base and contact profiles in the app's private storage area to enable context-aware completions.