X Recommender

## What is X‑Recommender?

X‑Recommender is an AI-powered browser extension that analyzes and scores tweets in your X (Twitter) feed. Using large language models (LLMs), it evaluates tweet quality and dynamically adjusts transparency — so high‑quality content stays visible while low‑quality content fades into the background.

The extension analyzes publicly available, user‑generated content on X and helps you personalize your browsing experience.

---

## Why You Need This

Tired of your X feed being flooded with:

- 🗣️ Low-value drama and gossip
- 📢 Spam and excessive self-promotion
- 💰 Crypto scams and clickbait
- 🤬 Toxic or unproductive content

**X‑Recommender filters out the noise so you can focus on meaningful content.**

---

## Key Features

### 🤖 AI-Powered Scoring
- Uses local or cloud-based LLM providers to analyze tweets
- Scores range from 0–100 based on information value, originality, clarity, and relevance
- Customizable prompts to match your personal preferences

### 👁️ Smart Transparency
- Low-scoring tweets automatically become transparent
- High-quality content remains fully visible
- Adjustable threshold for flexible filtering

### ⚙️ Flexible Model Support
Supports popular LLM providers, giving you the flexibility to use local models or cloud-based APIs.

### 🔒 Privacy & Control
- When using local models, data processing happens on your device
- If you choose a cloud API, tweet content is sent only to the provider you configure
- No tweet data is collected or stored by the extension developer

### 🔄 Smart Caching
- Remembers previously analyzed tweets
- Reduces repeated API calls
- Improves performance and efficiency

---

## How It Works

1. Install the extension
2. Configure your preferred LLM provider (local or cloud)
3. Browse X as usual
4. Tweets are analyzed automatically in the background
5. Low-quality tweets fade out while valuable content remains clear
6. Hover over a tweet to view its AI score and evaluation

---

## Important Notes

- This extension analyzes publicly available user-generated content on X.
- X and Twitter are trademarks of X Corp.
- This extension is not affiliated with, endorsed by, or sponsored by X Corp.

## 1.5 Updates
### New Features
- **ID Whitelist**: Add trusted users to a whitelist - their tweets will not be filtered and display as ∞
- **Filter Mode**: Choose between opacity (fade) or hide completely
- **No-Text Tweet Scoring**: Option to score tweets with only images/videos
- **Feedback Link**: Added feedback and follow link in the footer
- **Three-Tab Settings**: Reorganized settings into API, Filter, and Whitelist tabs

### Improvements
- Adjusted retweet badge position

## Version 1.5.1
### Fixes
- Fixed hide mode filtering to only use the hide threshold, preventing overlap with opacity behavior
- Improved tweet text fallback extraction (`innerText || textContent`) to reduce unstable hide/show flicker

## Version 1.6.0
### New Features
- Added support for multiple models - you can now configure and switch between multiple LLM APIs
- Automatically disable thinking mode with model that support thinking

### Fixes
- Fixed issue where adding a second model would not display in the list (race condition and storage sync delay)

## Version 1.6.1
### Improvements
- Pre-hide unscored tweets: New tweets stay hidden until scored, then are shown or hidden by score (reduces timeline jumps)
- Changed batch size from 5 to 1 to lower timeline jump probability
- Changed temperature from 0.1 to 0 for more stable scoring
- Added cache check before scoring to avoid duplicate requests
- Added viewport detection to prevent large-scale node removal outside viewport causing scroll anchor jumps
- Optimized tweet ID extraction logic

## Version 1.6.2
### Improvements
- Thread pass-through optimization: tweets in the same conversation now reuse the first scored result, avoiding repeated scoring for follow-up tweets
- Improved conversation detection by prioritizing `conversation_id` grouping (with safe fallback logic), covering mixed-author thread cases like repost + comment chains
- Reduced scroll flicker by disabling opacity transition on cached re-apply paths during timeline refresh/scroll
- Moved adaptive viewport timeout controls under hide mode only; these options are now shown only when filter mode is set to hide

## Version 1.6.3
### Fix
- Fix issue where preset APIs like Ollama couldn't be added after changing URL to LAN IP (preset values overriding user input)

### Improvements
- Preset API name and URL now prioritize user input

## Version 1.6.4
### Fixes
- Improved LM Studio compatibility for Qwen-style models by making the no-thinking fallback path more robust
- Prevented score parsing from accidentally reading reasoning step numbers as final scores

### Improvements
- Added an LM Studio setup hint recommending non-thinking models first
- Added guidance for advanced users to disable thinking via LM Studio Prompt Template settings

## Version 1.7.0
### New Features
- Added "Clear Cache" button to manually clear all cached scores