Question 1

How to fix 'Inference embeddings response contains an invalid vector' error when using LM Studio with RSSMonster?

Accepted Answer

When using `EMBEDDING_PROVIDER=openai-compatible` with LM Studio, the `openai` npm SDK (v7.10.0) defaults to requesting `encoding_format: 'base64'`. LM Studio ignores this parameter and always returns a plain JSON float array, which the SDK mis-decodes into a corrupted/truncated vector (e.g., 192 values instead of 768, with leading zeros). The server then rejects it with `Inference embeddings response contains an invalid vector`. Fix (confirmed working): explicitly force float format in `inference/src/embeddings/providers/openaiEmbeddingProvider.js`:

```js
const response = await client.embeddings.create({
  model: config.modelId,
  input: texts,
  encoding_format: 'float'
});
```

Alternatively, update RSSMonster to the latest version — the maintainer committed a patch that resolves this, verified by the reporter.

Question 2

Why does OIDC login silently redirect back to the login page without completing account linking?

Accepted Answer

The PWA service worker is intercepting the OIDC callback. The IdP redirect to `/api/auth/oidc/callback` is a navigation request, so Workbox's NavigationRoute serves the cached `index.html` instead of letting the request reach the server — the callback never appears in server logs and account linking silently fails. Fix: add `/^\/api\//` to `NAVIGATION_FALLBACK_DENYLIST` in `client/pwa-policy.js` so API routes bypass the navigation fallback. This was fixed upstream, so updating to a newer image also resolves it. To verify the diagnosis before patching, enable 'Bypass for network' in Chrome DevTools (Application > Service Workers) — if login then completes, the service worker is the cause.

Question 3

How to fix SEMANTIC_LABEL_INVALID_RESULT and empty AI responses in RSSMonster when using reasoning/thinking models (LM Studio, Ollama, Qwen3, DeepSeek-R1, Gemma)?

Accepted Answer

Cause: reasoning-capable models spend the entire max_tokens budget on their internal `reasoning_content` chain-of-thought and return empty `content` with `finish_reason: "length"`, so RSSMonster's structured-output jobs (semantic labels, article scoring/classification, smart-folder recommendations, feed rediscovery) fail silently with errors like `SEMANTIC_LABEL_INVALID_RESULT` or `CompletionBudgetError` / `INFERENCE_COMPLETION_BUDGET_EXHAUSTED`. Fix (per-request, no need to change model-level settings in LM Studio/Ollama): add these variables to `inference/.env` and restart the inference service:

GENERATION_REASONING_EFFORT=none
CLASSIFICATION_REASONING_EFFORT=none

This sends `reasoning_effort: 'none'` with each chat completion request, disabling thinking only for RSSMonster's calls while leaving the model's global reasoning behavior untouched for other apps. Note: `enable_thinking: false` may not work depending on backend/template — `reasoning_effort: 'none'` is the reliable option. Also disable speculative decoding (draft model) if results are still oversized or malformed.

Question 4

How to fix 'Network Error' (ERR_CONNECTION_REFUSED) on login when running behind a reverse proxy like Cloudflare Tunnel?

Accepted Answer

Cause: after an image upgrade, the frontend JavaScript bundle sent login requests to `http://localhost:3000/api/auth/login` instead of your public FQDN, so the browser blocked it with `POST http://localhost:3000/api/auth/login net::ERR_CONNECTION_REFUSED` when accessing via a reverse proxy (cloudflared, nginx, etc.). Fix: 1) Update to the latest image containing the patch: `docker compose pull && docker compose up -d`. 2) Clear the browser cache or hard-refresh the login page (Ctrl+Shift+R / Cmd+Shift+R) so the old cached JS bundle is replaced — without this step the error persists even after upgrading. Note: `PUBLIC_APP_URL` is only used for SMTP links and does not affect the login API URL.

Question 5

Why are RSSMonster sidebar article/unread counts stuck at 0 even though the articles exist?

Accepted Answer

This is a display issue, not data loss — the articles are intact in the database. The client loads the sidebar in two steps: structure first with 0 placeholders (`fetchOverviewLite`), then real numbers via `fetchOverviewCounts`. If the counts request is slow or fails (e.g., due to `SQLITE_BUSY: database is locked` contention on the SQLite Docker profile), the UI sets `overviewCountsStatus = 'error'` and stops — there is no automatic retry, so all counters stay at 0 until manually refreshed. Workarounds: 1) Click the 'Counts could not refresh / Retry' banner that appears in the sidebar, or 2) simply reload the page — the correct counts will return. To confirm it's this bug and not a filtering issue, check the SQLite DB directly: affected articles should have `aiAnalysisStatus='complete'`, `filteredInd=0`, and your account's `minAdvertisementScore`/`minSentimentScore`/`minQualityScore` thresholds should be 0. To reduce occurrences, make sure you're on the latest master which includes WAL mode and `busy_timeout=5000` in `server/config/databaseRuntime.js` (mitigates but does not fully eliminate SQLite locking on low-RAM NAS setups). An automatic bounded retry/backoff for `fetchOverviewCounts` has been requested but is not yet implemented.

RSSMonster

Core Features

What It Can't Do

Use Cases

Tags

Getting Started

Similar Projects