📡 RSS RAG Cache Plugin for LM Studio

Turn LM Studio into your own private, intelligent news archive. This plugin fetches your favorite RSS feeds, chunks the articles, embeds them using your local LM Studio server, and stores them in a local SQLite database for Retrieval-Augmented Generation (RAG).

Instead of searching the web, your LLM can instantly query your local cache of recent articles, blog posts, and newsletters.

✨ Features

• Local RAG Pipeline: Fetches RSS feeds, chunks text, and generates embeddings via LM Studio's local server.
• Hybrid Search (BM25 + RRF): Combines two SQLite FTS5 indexes (trigram for CJK/substrings, and porter unicode61 for real BM25 word ranking) with a resident in-memory vector index (falling back to a bounded SQL scan when the matrix would exceed the memory ceiling), fused by Reciprocal Rank Fusion. Active filters (feed / author / tag / collection / date range) are resolved to an entry-ID allow-list before ranking, so both keyword lists and the vector scan only score in-scope articles — a filtered search on a large cache costs time proportional to the matches, not to the cache. If the embedding model is offline, it automatically falls back to pure keyword search.
• Contextual Chunk Embeddings: Each chunk is embedded as title | section breadcrumb + newline + chunk (embed style contextual-v2), so summary-only feeds and mid-article passages still produce self-contained vectors. Document/query inputs then receive model-aware asymmetric prefixes (search_document: / search_query: for nomic-embed-text-*, Represent this sentence… for bge/mxbai/snowflake). The stored/displayed text is unchanged. Switchable via contextualChunks; caches on older embed styles are upgraded lazily by reindex_missing_embeddings.
• Heading Breadcrumbs: <h1>–<h3> are preserved as Markdown headings during extraction, and each chunk records the heading path that was active where the chunk ends (e.g. Title > Setup > Windows), so expanded context tells the model where a passage came from.
• Small-to-Big Context: Each search hit is expanded to include up to contextChunks neighbouring chunks, so the returned snippet carries surrounding context rather than an isolated passage.
• Fast, Incremental Vector Index: All embeddings are packed once into a single, densely-packed, pre-normalised Float32Array, so a semantic query is one tight dot-product loop with zero per-query allocation. The index is then maintained in place: newly indexed chunks append into recycled slots (tombstones tracked through an O(1) free list) and removed articles are tombstoned, so refreshes and retention no longer rebuild the whole matrix. It still rebuilds on a dimension change or when tombstones exceed 20% of capacity; a configurable byte ceiling (RSS_RAG_INDEX_MAX_MB, default 256 MB, with 15% headroom) falls back to the streaming scan when exceeded.
• Feed Auto-Discovery: manage_rss_feeds accepts a blog homepage and follows <link rel="alternate" type="application/rss+xml|atom+xml"> to the real feed, adding that instead. The granular test_feed reports any discovered alternates.
• Per-Feed TTL & Collections: feeds.json supports a per-feed retentionDays override and a collection label; retention honours the override, and search can be scoped with the collection filter.
• Agent-Optimized Tool Surface: The default surface is just 4 tools — search_rss_cache, get_rss_article, get_recent_cache_articles, manage_rss_feeds — so every chat only pays the token cost of 4 tool schemas. One multiplexed admin_rss_cache tool (gated by enableAdminTools) covers diagnostics/optimization/reindexing/debug, and the legacy one-purpose tools come back with exposeGranularAdminTools.
• Full Article Reader: get_rss_article returns a cached article's complete text and metadata given an entryId or canonical url, with offset/maxChars paging (default 12,000 chars) and an optional chunk layout (includeChunks: true) — so the model can read past the 600-character search snippet.
• Ranking Debug: admin_rss_cache with action: "debug_search" (or the granular search_rss_cache_debug) explains each result — per-list rank, RRF contribution, cosine similarity, fused score — for observability without changing behaviour.
• Eval Harness: npm run eval indexes a frozen corpus against a deterministic fake embedding server and reports recall@1/3/5/10, MRR, nDCG@10, and latency p50/p95, gating on regressions versus a committed baseline.
• Structure-Aware Chunking: Text is split on paragraph and sentence boundaries (with a small overlap) instead of a fixed character window, so chunks never cut a sentence in half and embeddings represent coherent passages.
• Content-Change Re-Indexing: Each entry stores a content hash; when a feed edits an article (same GUID, new body), the entry's chunks are rebuilt and re-embedded instead of being ignored forever.
• Robust Article Extraction: Reads <article> / content selectors, strips boilerplate, preserves paragraph breaks and heading structure, and falls back to Open Graph / meta descriptions.
• Self-Healing Embeddings: A bounded backfill runs once the model is detected at startup and after every refresh, so gaps from an offline model repair themselves. Embedding batches run with bounded concurrency and repeated queries are cached.
• Per-Feed Configuration (feeds.json): Optionally override enabled, tags, scrapeFullText, linkPatterns, maxItems, and request headers for individual feeds; tags become a search filter.
• Near-Duplicate Collapse (64-bit SimHash): Each entry stores a 64-bit SimHash over 3-gram word shingles (FNV-1a); syndicated articles that differ only in a footer or byline collapse to a single result, while genuinely distinct short updates are left uncollapsed.
• Rich Search Controls: Filter by days/after/before, feed, author, or tag; sort by relevance or date; paginate with limit/offset. Filters are applied before ranking, not after.
• Diagnostics: cache_diagnostics (or admin_rss_cache with action: "diagnostics") reports coverage, sizes, feed health, and suggested fixes.
• Conditional Fetching: Background refreshes send If-None-Match / If-Modified-Since (ETag / Last-Modified) so unchanged feeds answer 304 Not Modified and cost almost no bandwidth. A manual refresh (or force: true) re-downloads everything.
• Retry & Per-Feed Backoff: Requests retry on 429/5xx (and on a 403 that carries Retry-After) using the header (capped at 15s). Feeds that keep failing are placed in an exponential backoff (15 min doubling to 24 h) and skipped until their retry time, so one broken feed can't slow every refresh.
• Embedding-Model Guard: The model ID and vector dimension are recorded with the cache. If you switch models, a dimension mismatch is detected and surfaced (instead of silently degrading search) and admin_rss_cache with action: "reembed" rebuilds the index.
• Diversified, Recency-Aware Ranking: Results blend the fused relevance score with a gentle recency boost and pass through MMR (Maximal Marginal Relevance) so syndicated near-duplicate headlines don't crowd out distinct results. Snippets are windowed around the matching query terms.
• Auto-HTML Scraper (Cheerio): If a URL doesn't have an RSS feed (like a standard news webpage), the plugin automatically scrapes the page for article links using configurable keywords (news, article, post, ...).
• Smart Throttling (per host): Fetches RSS feeds concurrently across different domains, but never runs more than maxConcurrentPerHost requests at once against the same host and spaces them by minRequestGapMs + jitter (1.5–3 s by default). HTML pages are scraped one at a time, and robots.txt is respected before scraping or fetching full article text.
• Ban Detection & Auto-Pause: Cloudflare / Akamai / Incapsula / captcha walls (including JS challenges served with HTTP 200) are recognised, logged as a loud ⚠️ BAN warning, and the offending host is paused (30 min, doubling to 12 h). Paused hosts appear as blockedHosts in the refresh result, so a block is never silent.
• RSS/Atom XML via fast-xml-parser: Robust parsing of RSS 2.0 and Atom feeds, with a regex-based fallback parser for malformed XML.
• Admin Tools (Opt-In): Feed add/remove/test/configure, pruning, reindexing, and VACUUM are available from chat through manage_rss_feeds and admin_rss_cache; the legacy granular tools remain available behind exposeGranularAdminTools.
• Configurable Retention: Keeps entries by publish date (falling back to cache date) for 30, 60, 90, or 365 days.
• Startup Model Check: Warns (non-fatally) if your configured embedding model isn't loaded, so you know why search drops to keyword-only.
• Native SQLite Database: Uses Node.js's built-in node:sqlite module (WAL mode). No C++ addons or Visual Studio build tools required!
• Versioned Migrations: The schema is upgraded through an idempotent PRAGMA user_version migration runner (currently version 11), so new columns/indexes and data upgrades (like the 64-bit SimHash backfill) are applied safely to existing databases.
• Resilient Indexing: Articles and their chunks are embedded first and committed in a single transaction, so an interrupted or failed refresh never leaves half-indexed entries behind. Store-time L2 norms keep semantic search fast.
• Automatic Maintenance: After each refresh the retention pass runs PRAGMA optimize and VACUUMs the database only when enough free pages have accumulated (so routine refreshes stay fast).
• Safe Data Storage: Stores your feeds.txt and cache.db in ~/.lmstudio/plugin-data/rss-rag-cache/ so they survive plugin updates.

📋 Requirements

• LM Studio: Running the local server.
• Embedding Model: You must download and load an embedding model in LM Studio (e.g., nomic-embed-text-v1.5) and start the local server on port 1234.
• Node.js v22+: Requires Node version 22 or higher for native SQLite support.

🚀 Installation

• Download or install this plugin directly from the LM Studio Plugin Hub.
• Ensure LM Studio is running an embedding model on the Local Server.
• The plugin will automatically create feeds.txt in your plugin-data folder on the first run.

⚙️ Configuration

1. Adding Feeds (feeds.txt)

Your feeds file is located at:

• Windows: C:\Users\YOUR_USERNAME\.lmstudio\plugin-data\rss-rag-cache\feeds.txt
• Mac/Linux: ~/.lmstudio/plugin-data/rss-rag-cache/feeds.txt

Add one RSS/Atom URL per line. You can also add standard webpage URLs (the plugin will scrape them automatically). Lines starting with # are ignored.

2. Plugin Settings UI

Cache and indexing settings are global (they configure the shared on-disk cache). Only topK is per-chat, since it governs how much context a single conversation pulls in.

Global plugin settings:

Setting	Default	Description
retentionDays	90	How long to keep articles (30, 60, 90, or 365 days).
refreshIntervalHours	6	How often the plugin automatically fetches new RSS items in the background.
embeddingModel	nomic-embed-text-v1.5	The exact model ID loaded in your LM Studio server.
chunkSize	800	Length (in characters) of each text chunk used for embedding and search.
embedBatchSize	16	How many chunks are sent to the embedding model per request (higher = faster but more VRAM).
embedConcurrency	3	How many embedding requests run in parallel during indexing/backfill.
maxConcurrentFeeds	4	How many RSS/Atom feeds are fetched at the same time (across all hosts).
maxConcurrentPerHost	2	Cap on simultaneous requests to any single domain. Low values keep anti-bot firewalls calm.
minRequestGapMs	1500	Minimum delay between two requests to the same host. Higher = safer, slower.
requestGapJitterMs	1500	Random extra delay added to each per-host gap so the crawl pattern is not perfectly regular.
respectRobotsTxt	On	Check robots.txt before scraping HTML pages or fetching full article text.

Per-chat settings:

Setting	Default	Description
topK	8	How many unique articles to return per search.

3. Per-Feed Options (feeds.json)

For fine-grained control, create feeds.json next to feeds.txt (or use the set_feed_options tool). Feeds listed in feeds.txt are merged with their JSON overrides:

Supported fields per feed: enabled, tags, scrapeFullText, linkPatterns, maxItems, headers, retentionDays, collection. Set enabled: false to keep a feed configured but skip fetching it, and use tags and collection to filter searches.

🛠️ LM Studio Tools

The default surface is intentionally small — 4 tools — so each chat only pays the token cost of 4 tool schemas.

Default tools (always registered)

Optional: enableAdminTools

Optional: exposeGranularAdminTools (adds admin_rss_cache plus the legacy one-purpose tools)

🧪 Development

Contributor notes:

🧠 How It Works

🐛 Troubleshooting

Search returns 0 matches Ensure your embedding model is loaded and the server is running. If the model is offline, the plugin falls back to keyword search — but if your articles don't contain the exact keywords, they won't be found. Run manage_rss_feeds with action: "list" to check that chunks > 0; with enableAdminTools on, admin_rss_cache with action: "reindex" fills the gaps once the model is loaded.

HTTP 403 Forbidden on a feed Some websites block automated requests. The plugin uses a Chrome User-Agent, robots.txt, per-host throttling, and never retries a hard 403 — but strict firewalls might still block it. When that happens the host is paused automatically and a ⚠️ BAN warning names it, with the paused hosts listed under blockedHosts in the refresh result. If you control the site, set contactUserAgent so they can whitelist you; otherwise use an alternative RSS URL or an RSS Bridge service.

HTTP 429 Too Many Requests Sites like Reddit aggressively block Node.js fetch requests. If this happens, use an RSS Bridge URL (e.g., https://rss-bridge.org/bridge01/?action=display&bridge=RedditBridge&format=Atom) instead of the native .rss link.

I changed the embedding model and search got worse Vectors from different models aren't comparable. The plugin logs a dimension-mismatch / model-change warning and manage_rss_feeds (action: "list") shows the model and dimension the cache was built with. Run admin_rss_cache with action: "reembed" (enable enableAdminTools, or use reembed_all when granular tools are exposed) with the new model loaded to rebuild the index.

A feed keeps failing Repeated failures trigger exponential backoff (15 min → 24 h); manage_rss_feeds with action: "list" shows each feed's failure count and next retry time. Fix the URL, or run manage_rss_feeds with action: "refresh" and force: true to retry immediately.

📦 Changelog

1.5.0

1.4.0

1.3.0

1.2.0

1.1.0

1.0.0

• Initial public release.

📡 RSS RAG Cache Plugin for LM Studio

Turn LM Studio into your own private, intelligent news archive. This plugin fetches your favorite RSS feeds, chunks the articles, embeds them using your local LM Studio server, and stores them in a local SQLite database for Retrieval-Augmented Generation (RAG).

Instead of searching the web, your LLM can instantly query your local cache of recent articles, blog posts, and newsletters.

✨ Features

• Local RAG Pipeline: Fetches RSS feeds, chunks text, and generates embeddings via LM Studio's local server.
• Hybrid Search (BM25 + RRF): Combines two SQLite FTS5 indexes (trigram for CJK/substrings, and porter unicode61 for real BM25 word ranking) with a resident in-memory vector index (falling back to a bounded SQL scan when the matrix would exceed the memory ceiling), fused by Reciprocal Rank Fusion. Active filters (feed / author / tag / collection / date range) are resolved to an entry-ID allow-list before ranking, so both keyword lists and the vector scan only score in-scope articles — a filtered search on a large cache costs time proportional to the matches, not to the cache. If the embedding model is offline, it automatically falls back to pure keyword search.
• Contextual Chunk Embeddings: Each chunk is embedded as title | section breadcrumb + newline + chunk (embed style contextual-v2), so summary-only feeds and mid-article passages still produce self-contained vectors. Document/query inputs then receive model-aware asymmetric prefixes (search_document: / search_query: for nomic-embed-text-*, Represent this sentence… for bge/mxbai/snowflake). The stored/displayed text is unchanged. Switchable via contextualChunks; caches on older embed styles are upgraded lazily by reindex_missing_embeddings.
• Heading Breadcrumbs: <h1>–<h3> are preserved as Markdown headings during extraction, and each chunk records the heading path that was active where the chunk ends (e.g. Title > Setup > Windows), so expanded context tells the model where a passage came from.
• Small-to-Big Context: Each search hit is expanded to include up to contextChunks neighbouring chunks, so the returned snippet carries surrounding context rather than an isolated passage.
• Fast, Incremental Vector Index: All embeddings are packed once into a single, densely-packed, pre-normalised Float32Array, so a semantic query is one tight dot-product loop with zero per-query allocation. The index is then maintained in place: newly indexed chunks append into recycled slots (tombstones tracked through an O(1) free list) and removed articles are tombstoned, so refreshes and retention no longer rebuild the whole matrix. It still rebuilds on a dimension change or when tombstones exceed 20% of capacity; a configurable byte ceiling (RSS_RAG_INDEX_MAX_MB, default 256 MB, with 15% headroom) falls back to the streaming scan when exceeded.
• Feed Auto-Discovery: manage_rss_feeds accepts a blog homepage and follows <link rel="alternate" type="application/rss+xml|atom+xml"> to the real feed, adding that instead. The granular test_feed reports any discovered alternates.
• Per-Feed TTL & Collections: feeds.json supports a per-feed retentionDays override and a collection label; retention honours the override, and search can be scoped with the collection filter.
• Agent-Optimized Tool Surface: The default surface is just 4 tools — search_rss_cache, get_rss_article, get_recent_cache_articles, manage_rss_feeds — so every chat only pays the token cost of 4 tool schemas. One multiplexed admin_rss_cache tool (gated by enableAdminTools) covers diagnostics/optimization/reindexing/debug, and the legacy one-purpose tools come back with exposeGranularAdminTools.
• Full Article Reader: get_rss_article returns a cached article's complete text and metadata given an entryId or canonical url, with offset/maxChars paging (default 12,000 chars) and an optional chunk layout (includeChunks: true) — so the model can read past the 600-character search snippet.
• Ranking Debug: admin_rss_cache with action: "debug_search" (or the granular search_rss_cache_debug) explains each result — per-list rank, RRF contribution, cosine similarity, fused score — for observability without changing behaviour.
• Eval Harness: npm run eval indexes a frozen corpus against a deterministic fake embedding server and reports recall@1/3/5/10, MRR, nDCG@10, and latency p50/p95, gating on regressions versus a committed baseline.
• Structure-Aware Chunking: Text is split on paragraph and sentence boundaries (with a small overlap) instead of a fixed character window, so chunks never cut a sentence in half and embeddings represent coherent passages.
• Content-Change Re-Indexing: Each entry stores a content hash; when a feed edits an article (same GUID, new body), the entry's chunks are rebuilt and re-embedded instead of being ignored forever.
• Robust Article Extraction: Reads <article> / content selectors, strips boilerplate, preserves paragraph breaks and heading structure, and falls back to Open Graph / meta descriptions.
• Self-Healing Embeddings: A bounded backfill runs once the model is detected at startup and after every refresh, so gaps from an offline model repair themselves. Embedding batches run with bounded concurrency and repeated queries are cached.
• Per-Feed Configuration (feeds.json): Optionally override enabled, tags, scrapeFullText, linkPatterns, maxItems, and request headers for individual feeds; tags become a search filter.
• Near-Duplicate Collapse (64-bit SimHash): Each entry stores a 64-bit SimHash over 3-gram word shingles (FNV-1a); syndicated articles that differ only in a footer or byline collapse to a single result, while genuinely distinct short updates are left uncollapsed.
• Rich Search Controls: Filter by days/after/before, feed, author, or tag; sort by relevance or date; paginate with limit/offset. Filters are applied before ranking, not after.
• Diagnostics: cache_diagnostics (or admin_rss_cache with action: "diagnostics") reports coverage, sizes, feed health, and suggested fixes.
• Conditional Fetching: Background refreshes send If-None-Match / If-Modified-Since (ETag / Last-Modified) so unchanged feeds answer 304 Not Modified and cost almost no bandwidth. A manual refresh (or force: true) re-downloads everything.
• Retry & Per-Feed Backoff: Requests retry on 429/5xx (and on a 403 that carries Retry-After) using the header (capped at 15s). Feeds that keep failing are placed in an exponential backoff (15 min doubling to 24 h) and skipped until their retry time, so one broken feed can't slow every refresh.
• Embedding-Model Guard: The model ID and vector dimension are recorded with the cache. If you switch models, a dimension mismatch is detected and surfaced (instead of silently degrading search) and admin_rss_cache with action: "reembed" rebuilds the index.
• Diversified, Recency-Aware Ranking: Results blend the fused relevance score with a gentle recency boost and pass through MMR (Maximal Marginal Relevance) so syndicated near-duplicate headlines don't crowd out distinct results. Snippets are windowed around the matching query terms.
• Auto-HTML Scraper (Cheerio): If a URL doesn't have an RSS feed (like a standard news webpage), the plugin automatically scrapes the page for article links using configurable keywords (news, article, post, ...).
• Smart Throttling (per host): Fetches RSS feeds concurrently across different domains, but never runs more than maxConcurrentPerHost requests at once against the same host and spaces them by minRequestGapMs + jitter (1.5–3 s by default). HTML pages are scraped one at a time, and robots.txt is respected before scraping or fetching full article text.
• Ban Detection & Auto-Pause: Cloudflare / Akamai / Incapsula / captcha walls (including JS challenges served with HTTP 200) are recognised, logged as a loud ⚠️ BAN warning, and the offending host is paused (30 min, doubling to 12 h). Paused hosts appear as blockedHosts in the refresh result, so a block is never silent.
• RSS/Atom XML via fast-xml-parser: Robust parsing of RSS 2.0 and Atom feeds, with a regex-based fallback parser for malformed XML.
• Admin Tools (Opt-In): Feed add/remove/test/configure, pruning, reindexing, and VACUUM are available from chat through manage_rss_feeds and admin_rss_cache; the legacy granular tools remain available behind exposeGranularAdminTools.
• Configurable Retention: Keeps entries by publish date (falling back to cache date) for 30, 60, 90, or 365 days.
• Startup Model Check: Warns (non-fatally) if your configured embedding model isn't loaded, so you know why search drops to keyword-only.
• Native SQLite Database: Uses Node.js's built-in node:sqlite module (WAL mode). No C++ addons or Visual Studio build tools required!
• Versioned Migrations: The schema is upgraded through an idempotent PRAGMA user_version migration runner (currently version 11), so new columns/indexes and data upgrades (like the 64-bit SimHash backfill) are applied safely to existing databases.
• Resilient Indexing: Articles and their chunks are embedded first and committed in a single transaction, so an interrupted or failed refresh never leaves half-indexed entries behind. Store-time L2 norms keep semantic search fast.
• Automatic Maintenance: After each refresh the retention pass runs PRAGMA optimize and VACUUMs the database only when enough free pages have accumulated (so routine refreshes stay fast).
• Safe Data Storage: Stores your feeds.txt and cache.db in ~/.lmstudio/plugin-data/rss-rag-cache/ so they survive plugin updates.

📋 Requirements

• LM Studio: Running the local server.
• Embedding Model: You must download and load an embedding model in LM Studio (e.g., nomic-embed-text-v1.5) and start the local server on port 1234.
• Node.js v22+: Requires Node version 22 or higher for native SQLite support.

🚀 Installation

• Download or install this plugin directly from the LM Studio Plugin Hub.
• Ensure LM Studio is running an embedding model on the Local Server.
• The plugin will automatically create feeds.txt in your plugin-data folder on the first run.

⚙️ Configuration

1. Adding Feeds (feeds.txt)

Your feeds file is located at:

• Windows: C:\Users\YOUR_USERNAME\.lmstudio\plugin-data\rss-rag-cache\feeds.txt
• Mac/Linux: ~/.lmstudio/plugin-data/rss-rag-cache/feeds.txt

Add one RSS/Atom URL per line. You can also add standard webpage URLs (the plugin will scrape them automatically). Lines starting with # are ignored.

2. Plugin Settings UI

Cache and indexing settings are global (they configure the shared on-disk cache). Only topK is per-chat, since it governs how much context a single conversation pulls in.

Global plugin settings:

Setting	Default	Description
retentionDays	90	How long to keep articles (30, 60, 90, or 365 days).
refreshIntervalHours	6	How often the plugin automatically fetches new RSS items in the background.
embeddingModel	nomic-embed-text-v1.5	The exact model ID loaded in your LM Studio server.
chunkSize	800	Length (in characters) of each text chunk used for embedding and search.
embedBatchSize	16	How many chunks are sent to the embedding model per request (higher = faster but more VRAM).
embedConcurrency	3	How many embedding requests run in parallel during indexing/backfill.
maxConcurrentFeeds	4	How many RSS/Atom feeds are fetched at the same time (across all hosts).
maxConcurrentPerHost	2	Cap on simultaneous requests to any single domain. Low values keep anti-bot firewalls calm.
minRequestGapMs	1500	Minimum delay between two requests to the same host. Higher = safer, slower.
requestGapJitterMs	1500	Random extra delay added to each per-host gap so the crawl pattern is not perfectly regular.
respectRobotsTxt	On	Check robots.txt before scraping HTML pages or fetching full article text.

Per-chat settings:

Setting	Default	Description
topK	8	How many unique articles to return per search.

3. Per-Feed Options (feeds.json)

For fine-grained control, create feeds.json next to feeds.txt (or use the set_feed_options tool). Feeds listed in feeds.txt are merged with their JSON overrides:

Supported fields per feed: enabled, tags, scrapeFullText, linkPatterns, maxItems, headers, retentionDays, collection. Set enabled: false to keep a feed configured but skip fetching it, and use tags and collection to filter searches.

🛠️ LM Studio Tools

The default surface is intentionally small — 4 tools — so each chat only pays the token cost of 4 tool schemas.

Default tools (always registered)

Optional: enableAdminTools

Optional: exposeGranularAdminTools (adds admin_rss_cache plus the legacy one-purpose tools)

🧪 Development

Contributor notes:

🧠 How It Works

🐛 Troubleshooting

Search returns 0 matches Ensure your embedding model is loaded and the server is running. If the model is offline, the plugin falls back to keyword search — but if your articles don't contain the exact keywords, they won't be found. Run manage_rss_feeds with action: "list" to check that chunks > 0; with enableAdminTools on, admin_rss_cache with action: "reindex" fills the gaps once the model is loaded.

HTTP 403 Forbidden on a feed Some websites block automated requests. The plugin uses a Chrome User-Agent, robots.txt, per-host throttling, and never retries a hard 403 — but strict firewalls might still block it. When that happens the host is paused automatically and a ⚠️ BAN warning names it, with the paused hosts listed under blockedHosts in the refresh result. If you control the site, set contactUserAgent so they can whitelist you; otherwise use an alternative RSS URL or an RSS Bridge service.

HTTP 429 Too Many Requests Sites like Reddit aggressively block Node.js fetch requests. If this happens, use an RSS Bridge URL (e.g., https://rss-bridge.org/bridge01/?action=display&bridge=RedditBridge&format=Atom) instead of the native .rss link.

I changed the embedding model and search got worse Vectors from different models aren't comparable. The plugin logs a dimension-mismatch / model-change warning and manage_rss_feeds (action: "list") shows the model and dimension the cache was built with. Run admin_rss_cache with action: "reembed" (enable enableAdminTools, or use reembed_all when granular tools are exposed) with the new model loaded to rebuild the index.

A feed keeps failing Repeated failures trigger exponential backoff (15 min → 24 h); manage_rss_feeds with action: "list" shows each feed's failure count and next retry time. Fix the URL, or run manage_rss_feeds with action: "refresh" and force: true to retry immediately.

📦 Changelog

1.5.0

1.4.0

1.3.0

1.2.0

1.1.0

1.0.0

• Initial public release.