Aureliolo/steamgauge-search-encoder

SteamGauge search encoder

Qwen/Qwen3-Embedding-0.6B as a half-precision ONNX graph. SteamGauge uses it to search Steam reviews by meaning: the encoder finds the claims closest to a search, and the reranker orders them.

Use

• model.onnx takes input_ids and attention_mask from tokenizer.json, padded on the right.
• Returns vector: a normalised 1,024-dimension embedding per text, taken at the last token.
• Searches are written as Instruct: {instruction}\nQuery: {query}; the passages searched are embedded as they are.
• training/export_search.py measures its cosine similarity to the original model on DirectML, over any claims it is given.

Licence

Apache-2.0, as is Qwen/Qwen3-Embedding-0.6B. The model is Qwen's work.