System AtomCrafts

FeedCraft comes with a set of built-in “AtomCrafts” that perform specific processing steps on your feeds. You can chain these AtomCrafts together in a FlowCraft to create powerful pipelines.

Content enhancement

These atoms fetch full article content or clean up HTML.

fulltext

Extracts the full content of the article from the original webpage.

• Use case: When the RSS feed only provides a summary or snippet.
• Mechanism: Uses a standard HTTP client to fetch the page and an algorithm to extract the main content. Fast and lightweight.

fulltext-plus

Extracts full content using a headless browser (Puppeteer).

• Use case: For websites that require JavaScript to render content or have strong anti-bot protections.
• Mechanism: Connects to the configured browser provider (browserless-restful or cdp) to render the page. Slower but more robust.
• Parameters:
  • mode (default: networkidle2): Wait condition.
    • load: Wait for the load event.
    • domcontentloaded: Wait for the DOMContentLoaded event.
    • networkidle0: Wait until there are 0 active network connections for at least 500ms.
    • networkidle2: Wait until there are no more than 2 active network connections for at least 500ms. (Recommended for SPAs).
  • wait (default: 0): Explicit wait time in seconds (e.g., 5).

cleanup

Cleans up the HTML content to remove clutter.

• Use case: Improving readability by removing classes, styles, and empty tags.

relative-link-fix

Converts relative links (e.g., <a href="/about">) in the content to absolute links (e.g., <a href="https://example.com/about">).

• Use case: Essential when extracting full content, as relative links will break when viewed in an RSS reader.

guid-fix

Replaces the RSS item GUID with an FNV-1a hash of the item’s content.

• Use case: Some feeds change their GUIDs frequently even when content hasn’t changed, causing duplicate unread items in readers. This atom stabilizes the GUID based on content.

beautify-content

Re-formats the article using LLM to fix layout, remove ads, and standardizing Markdown, then converts back to clean HTML.

• Parameters:
  • prompt: Instructions for the “editor” persona.

---

AI processing

Use large language models to transform, enrich, or screen content.

Note

These atoms require LLM configuration (API Key, Base URL, etc.) in your environment variables.

translate-title

Translates the article title to your target language.

• Parameters:
  • prompt: Custom prompt. Defaults to a standard translation prompt. Supports {{.TargetLang}} placeholder.

translate-content

Translates the entire article content, replacing the original.

• Parameters:
  • prompt: Custom prompt. Supports {{.TargetLang}}.

translate-content-immersive

Bilingual translation. Appends the translation after each paragraph of the original text.

• Parameters:
  • prompt: Custom prompt.

summary

Generates a summary of the article and prepends it to the content.

• Parameters:
  • prompt: Custom prompt for summarization.

introduction

Generates a short introduction or “lead-in” for the article.

• Parameters:
  • prompt: Custom prompt.

ai-content-process

Process article content using LLM according to custom rules and insert the generated result at a specified position.

• Parameters:
  • rule (Required): Instruction for processing each article content. Example: “Summarize key points and list action items.”
  • extra-payload (Default: article_content): Comma-separated list of extra information to send to the LLM. Supported: article_summary, article_content, article_date, raw_rss_item.
  • placement (Default: prepend): Where to write the generated content. Supported: prepend, replace, append.

ignore-advertorial

Uses LLM to detect if an article is an advertorial or soft advertisement (evaluating both title and content) and removes it.

• Parameters:
  • prompt-for-exclude: A prompt that should return true if the item is an ad.

---

Filtering

Control which items remain in the feed, including keyword, time, and semantic filters.

keyword

Filters items based on keywords in the title or content.

• Parameters:
  • keywords: A comma-separated list of keywords to match (substring match, case-sensitive). Example: ad,sell,SALE.
  • mode: include (default) to keep matching items, or exclude to remove them.
  • scope: title, content, or all (default).

limit

Limits the number of items in the feed.

• Parameters:
  • num (default: 10): The maximum number of items to keep.

time-limit

Filters out items that are older than a specific number of days.

• Parameters:
  • days (default: 7): Max age of articles in days.

llm-filter

Generic LLM-based filter. You define the condition for exclusion. The LLM evaluates both the article title and content against this condition.

• Parameters:
  • filter_condition: A natural language question/condition. If the LLM answers “yes” (true), the item is removed.
  • Example: “Is this article about sports?” (Removes sports articles).

embedding-filter

Semantic topic filtering powered by an embedding model. Instead of asking a chat model to judge every item, FeedCraft converts both your topic anchors and each article into vectors, compares them with cosine similarity, and then keeps or removes matching items.

Tip

Use embedding-filter when you need fast, repeatable topic filtering such as “keep AI infrastructure news” or “remove sports articles”. Use llm-filter when the rule needs nuanced reasoning, policy interpretation, or structured decisions.

Environment variables

Set a dedicated embedding endpoint when possible:

FC_EMBEDDING_API_TYPE=openai
FC_EMBEDDING_API_BASE=https://api.openai.com/v1
FC_EMBEDDING_API_KEY=sk-your-api-key
FC_EMBEDDING_API_MODEL=text-embedding-3-small # Required
FC_EMBEDDING_BATCH_SIZE=5
FC_EMBEDDING_MAX_INPUT_CHARS=8000

Supported FC_EMBEDDING_API_TYPE values:

• openai: OpenAI or OpenAI-compatible embedding endpoint.
• gemini: Gemini through its OpenAI-compatible embedding endpoint. Set FC_EMBEDDING_API_BASE and FC_EMBEDDING_API_MODEL explicitly.
• ollama: Local Ollama embedding model. Set FC_EMBEDDING_API_BASE, for example http://localhost:11434, and an embedding model such as nomic-embed-text or bge-m3.

If FC_EMBEDDING_API_TYPE, FC_EMBEDDING_API_BASE, and FC_EMBEDDING_API_KEY are not set, FeedCraft falls back to the matching FC_LLM_API_TYPE, FC_LLM_API_BASE, and FC_LLM_API_KEY values. The embedding model name is independent: you must explicitly set FC_EMBEDDING_API_MODEL to a valid embedding model. FeedCraft does not provide a default embedding model and does not reuse FC_LLM_API_MODEL because that value is usually a chat model.

FC_EMBEDDING_MAX_INPUT_CHARS is a final safety cap applied to every text sent to the embedding service, including any instruction prefix. It is a character budget, not an exact tokenizer count. Keep it at or below a conservative value for your model’s token window, for example 8000 for an 8k-token embedding model.

Parameters

• anchors (required): One topic anchor per line. Each anchor should describe what you want to match, for example:

  artificial intelligence infrastructure
  machine learning research
  large language model deployment

• threshold (default: 0.6): Cosine similarity threshold from 0 to 1. Higher values are stricter. Start with 0.6, lower it if relevant articles are missing, and raise it if unrelated articles slip through.

• mode (default: include): include keeps matching items; exclude removes matching items.

• max_content_length (default: 2000): Maximum article content characters used by this AtomCraft before the final FC_EMBEDDING_MAX_INPUT_CHARS safety cap is applied.

• instruction (optional): Text prefix prepended to every embedding input. Leave it empty unless your model benefits from a fixed task prefix.

Admin UI workflow

1. Open Worktable → Process Feed Sources → AtomCraft.
2. Create a new AtomCraft, for example ai-news-only.
3. Select template embedding-filter.
4. Fill anchors with one topic per line.
5. Keep mode=include to keep matching items, or switch to exclude to remove matching items.
6. Save the AtomCraft.
7. Use it in a FlowCraft, Recipe, Feed Compare, or directly:

/craft/ai-news-only?input_url=https%3A%2F%2Fexample.com%2Ffeed.xml

Troubleshooting

• “anchors parameter is required”: Add at least one non-empty line to anchors.
• “FC_EMBEDDING_API_MODEL must be set”: Configure a single embedding model. Chat models are not suitable here.
• All items are removed: Lower threshold, add broader anchors, or increase max_content_length.
• Too many unrelated items remain: Raise threshold, make anchors more specific, or switch from broad category names to representative phrases.
• Provider says the input is too long: Lower FC_EMBEDDING_MAX_INPUT_CHARS, lower max_content_length, or shorten instruction.

---

Input & output

Pass-through proxy. HTML / JSON / Search to RSS converters are documented in Guides and System Tools.

proxy

Simple proxy for the feed.

• Use case: When you just want to forward the original feed without modification, or use FeedCraft as a central gateway.