Feed Digest Extension for FreshRSS

Automatically summarize newly retrieved RSS articles using LLM APIs (OpenAI-compatible). This extension processes articles during feed updates, creates combined summary articles in your destination language, and marks the originals as read.

Features

• 🤖 Automatic Summarization: Processes unread articles using LLM APIs during scheduled feed updates
• 🌍 Multi-language: Translates article titles and summaries to your chosen language
• ⚡ Efficient Batch Processing: Summarizes multiple articles in a single API call to reduce costs
• 📊 Per-feed Control: Enable/disable summarization and configure batch size for each feed individually
• 🎯 Smart Filtering: Skips image-only and too-short articles, adds explanatory notes
• 🎨 Clean Output: Creates formatted summary articles with links to originals

Requirements

• FreshRSS 1.24.0 or later
• PHP 7.4+ with cURL extension
• An OpenAI-compatible API key (OpenAI, Anthropic Claude, local models, etc.)
• Sufficient PHP max_execution_time (recommended: 300+ seconds for large batches)

Installation

1. Download or clone this repository
2. Copy the xExtension-FeedDigest directory to your FreshRSS extensions directory:

   cp -r xExtension-FeedDigest /path/to/FreshRSS/extensions/

3. In FreshRSS, navigate to Settings → Extensions
4. Enable the "Feed Digest" extension
5. Click "Configure" to set up your API credentials

Configuration

Global Settings

Navigate to Settings → Extensions → Feed Digest → Configure

Required settings:

• API Endpoint: OpenAI-compatible API endpoint URL

  • OpenAI: https://api.openai.com/v1
  • Anthropic Claude (via OpenAI compatibility): Check your provider's documentation
  • Local models: Your local endpoint URL

• API Secret Key: Your API authentication key

  • Keep this secure!
  • Never share or commit this key to version control

• Model Name: The LLM model to use

  • OpenAI: gpt-5.6-luna (recommended for current-generation, high-volume workloads), gpt-5.6-terra (higher quality)
  • Claude (through an OpenAI-compatible provider): claude-haiku-4-5 or the pinned claude-haiku-4-5-20251001
  • Other: Check your provider's model names

• Destination Language: Target language for summaries and translations

  • Examples: English, Spanish, Simplified Chinese, French, Japanese, German
  • The LLM will translate titles and write summaries in this language

• Max Content Length: Maximum characters per article (500-16000)

  • Default: 4000
  • Truncates longer articles to avoid LLM context limits
  • Estimate: 1 char ≈ 0.4 tokens

Per-Feed Settings

To enable summarization for a specific feed:

1. Navigate to Settings → Feeds
2. Select the feed you want to summarize
3. Scroll to the Feed Digest section
4. Configure the following:
   • Summarize articles with LLM: Set to Yes
   • Articles per summary batch: Number of articles to include in each summary (1-50, default: 10)
     • Articles are processed in batches to avoid timeouts
     • Each batch creates one summary article
     • Example: 35 unread articles with batch size 10 → 3 summary articles (10+10+10), 5 remain unread
5. Click Submit

API Endpoint Examples

OpenAI

Endpoint: https://api.openai.com/v1
Model: gpt-5.6-luna
Key: sk-...

Anthropic Claude (via OpenAI-compatible wrappers)

This extension currently uses OpenAI's Chat Completions request format, not Anthropic's native Messages API. To use Claude, configure an OpenAI-compatible provider or gateway and use its endpoint and model name. Claude Haiku 4.5 (claude-haiku-4-5) is a fast, cost-efficient option; verify the exact model identifier and pricing with your provider.

Local Models (Ollama, LM Studio, etc.)

Endpoint: http://localhost:11434/v1  # Ollama
Model: llama3.2
Key: not-needed  # Often not required for local models

OpenRouter

Endpoint: https://openrouter.ai/api/v1
Model: anthropic/claude-haiku-4.5
Key: sk-or-v1-...

How It Works

1. Scheduled Updates: During your regular FreshRSS cron/scheduled feed updates, the extension activates
2. Feed Check: For each feed with summarization enabled, it scans unread articles in pages and selects up to 200 eligible articles per maintenance run
3. Article Filtering:
   • Filters out previously created summary articles
   • Identifies image-only or too-short articles (< 100 characters)
   • Adds explanatory notes to skipped articles (they remain unread for you to review)
4. Batch Processing: Articles are processed in configurable batches (default: 10 per batch)
   • Only processes batches when enough articles are available
   • Each batch is sent to the LLM API in one request for efficiency
   • Each batch succeeds or fails independently
5. Summary Creation: For each batch, a new "summary" article is created with:
   • Translated titles (in your destination language)
   • Concise summaries (2-4 sentences each)
   • Links to original articles
   • Clean HTML formatting
6. Mark as Read: Only successfully summarized articles are marked as read
7. Auto-retry: Failed batches remain unread and will be retried on the next update

PHP Timeout Configuration

For large batches, you may need to increase PHP execution time:

In php.ini:

max_execution_time = 300

In FreshRSS .htaccess (Apache):

php_value max_execution_time 300

In Nginx config:

fastcgi_read_timeout 300;

Estimation:

• Each batch of 10 articles takes ~5-15 seconds (API call + processing)
• Multiple batches are processed sequentially per feed
• Recommended: 300 seconds (5 minutes) for safety with multiple feeds

Cost Estimation

API costs vary by provider, model, tokenization, and response length. The estimates below use standard list prices as of August 2026 and assume a typical batch of 10 articles at the default maximum of 4,000 characters each:

• Approximately 40,000 input characters, estimated as 16,000 input tokens
• Approximately 1,000 output tokens for 10 translated titles and summaries

Model	Input / 1M tokens	Output / 1M tokens	Estimated cost per batch
gpt-5.6-luna (recommended)	$1.00	$6.00	~$0.022
gpt-5.6-terra	$2.50	$15.00	~$0.055
Claude Haiku 4.5	$1.00	$5.00	~$0.021

Example scenario: 5 feeds, each with 20 unread articles/day, batch size 10:

• 5 feeds × 2 batches/day = 10 batches/day
• With gpt-5.6-luna: ~$0.22/day or ~$6.60/month
• With Claude Haiku 4.5 at Anthropic list prices: ~$0.21/day or ~$6.30/month

Actual usage is often lower because many articles are shorter than the configured maximum. Gateway pricing, reasoning tokens, cached tokens, retries, taxes, and provider-specific fees can change the total. Check the OpenAI model pricing and Anthropic model pricing before deployment.

Tip: Start with gpt-5.6-luna for current-generation quality at high volume. Claude Haiku 4.5 is a similarly priced alternative when used through an OpenAI-compatible provider.

Troubleshooting

API Connection Failed

1. Test your API connection using the "Test API Connection" button
2. Verify your API endpoint URL is correct
3. Check your API key is valid and has sufficient credits
4. Review FreshRSS logs for detailed error messages

Articles Not Being Summarized

1. Verify the feed has "Summarize articles with LLM" enabled
2. Check that articles are marked as unread
3. Ensure your API key is configured and valid
4. Look for errors in FreshRSS logs: data/users/_/log*.txt

PHP Timeout Errors

1. Increase max_execution_time in PHP configuration (recommended: 300 seconds)
2. Reduce "Articles per summary batch" setting for individual feeds
3. Disable summarization for some feeds to reduce total processing time

Summaries in Wrong Language

1. Check "Destination Language" setting is correct
2. Be specific (e.g., "Simplified Chinese" vs just "Chinese")
3. Test with a single article first

High API Costs

1. Use gpt-5.6-luna or Claude Haiku 4.5 for cost-sensitive workloads
2. Reduce "Articles per summary batch" for feeds (processes fewer articles at once)
3. Lower "Max Content Length" to send less data per article
4. Enable summarization only for high-value feeds
5. Monitor API usage on your provider's dashboard

Privacy & Data Usage

• API Calls: Article content is sent to your configured LLM API
• Data Storage: Only summaries are stored locally; API has its own data retention policies
• Security: API keys are stored in FreshRSS configuration (keep backups secure)
• Logging: Errors and processing info logged to FreshRSS logs

Limitations

• Cron-based: Summarization happens during scheduled updates, not immediately on manual refresh
• Batch Processing: Articles must accumulate to the configured batch size before processing
• Sequential Batches: Each feed's batches are processed sequentially to avoid timeouts
• No Retry Tracking: Failed batches retry every update (no exponential backoff)
• Context Limits: Very long articles are truncated based on max content length setting
• Image-only Articles: Articles with minimal text are skipped and left unread with an explanatory note

Development

Testing

To test the extension:

1. Enable for a single test feed with few articles
2. Manually trigger feed update
3. Check logs for processing messages
4. Verify summary article appears in feed
5. Confirm original articles marked as read

Debugging

Enable detailed logging in FreshRSS and monitor:

• data/users/_/log.txt or data/users/_/log_*.txt
• Look for "Feed Digest:" prefixed messages
• The feed inventory line shows every enabled feed with its ID, total article count, and cached unread count
• Large-feed scan lines show how many rows were scanned and how many were generated summaries, already processed, newly skipped, or eligible

Support

For issues, questions, or contributions:

• GitHub Issues: https://github.com/fengchang/xExtension-FeedDigest
• FreshRSS Community: https://github.com/FreshRSS/FreshRSS/discussions

License

GNU Affero General Public License v3.0 (AGPL-3.0)

See LICENSE file for details.

Credits

Developed for the FreshRSS community.

---

Note: This extension uses third-party AI services. Review their terms of service and privacy policies before use.