Qwen Embeddings

Historical evaluation of the architecture before Topic removal. The measurements below are retained for comparison and do not describe current processing.

RSSMonster can run onnx-community/Qwen3-Embedding-0.6B-ONNX locally through Transformers.js. It uses CPU inference, last-token pooling, L2 normalization, and the model’s native 1024-dimensional vectors. One model instance is loaded during service startup and reused, and inference requests are serialized in-process.

Configuration

Enable the server capability flags in Model Usage and configure its inference connection. A longer timeout is useful on low-power hardware and during the first model load:

# server/.env
INFERENCE_BASE_URL=http://127.0.0.1:3001
INFERENCE_TIMEOUT_MS=600000

Configure Qwen in the inference service:

# inference/.env
INFERENCE_HOST=127.0.0.1
INFERENCE_PORT=3001
INFERENCE_DEBUG=false
INFERENCE_MODEL_CACHE_DIR=.cache/models
EMBEDDING_PROVIDER=local
GENERATION_PROVIDER=local
GENERATION_MODEL=onnx-community/Qwen3.5-0.8B-ONNX
GENERATION_DTYPE=q4
CLASSIFICATION_PROVIDER=local
EMBEDDING_MODEL=onnx-community/Qwen3-Embedding-0.6B-ONNX
EMBEDDING_DIMENSIONS=1024
EMBEDDING_MAX_BATCH_SIZE=8
EMBEDDING_QUEUE_MAX_PENDING=4

Qwen can provide embeddings separately from Qwen3.5 generation. With the configuration above, Qwen3.5 generates article summaries, tags, Smart Folder recommendations, and feed rediscovery results. ModernBERT scores article quality locally. These capabilities require no OpenAI key. The optional assistant uses an independently configured compatible endpoint; follow Assistant and MCP to enable it separately.

Qwen embedding inference runs one batch at a time and accepts four pending batches by default. Set EMBEDDING_QUEUE_MAX_PENDING to a positive integer to adjust that bound. Pending work is removed when its HTTP caller disconnects; requests beyond the bound receive the inference queue’s stable 503 overload response.

Transformers.js downloads missing model assets while the inference service starts and stores them under inference/.cache/models by default. Relative cache paths are resolved from the inference project directory. Cached model files are not committed to Git and are reused after restarts. Deleting the cache causes the model to be downloaded again on its next use.

The HTTP listener opens before models finish loading. /health reports liveness, while /ready returns 503 until the configured models are ready. With INFERENCE_DEBUG=true (automatically enabled by npm run dev), the console reports loaded:true, readiness for crawling, and content-safe request and completion messages for embedding, generation, and other inference capabilities.

Observed Semantic Behavior

In RSSMonster’s 686-article semantic regression fixture, Qwen passed all 12 pipeline tests. Compared with OpenAI text-embedding-3-small, it grouped more articles into events and topics and separated two unrelated transport accidents that OpenAI placed in one topic. It also formed more complete groups for several news stories.

Qwen was more aggressive when assigning content through interest-island fallback paths, including some broad or questionable assignments, and it split a few article sets that OpenAI grouped together. Duplicate detection was the same for both models. This suggests stronger event/topic recall in the current fixture, with a greater need to calibrate island behavior. The outcome depends on the articles, prompts, and similarity thresholds and should not be treated as a universal ranking of the models.

Qwen keeps embedding content and inference on the host and avoids per-request API charges. In exchange, the host must store the model and provide sufficient memory and CPU time. Its vectors cannot be mixed with OpenAI vectors.