ecoai-rag-encoder
Fitted retriever for FindExpert.ir ecoAI (not a 14B LoRA, not a MiniLM fork).
This environment has ~2GB RAM and no GPU, so Qwen/E5 cannot be trained here. v2 = unicode word unigrams+bigrams, sublinear tf-idf, L2 cosine, extra Iran/grant recalls — still not a neural LoRA.
RAG Space that uses these weights: sosa123454321/ecoai-space
| File | Role |
|---|---|
encoder.joblib |
sklearn TfidfVectorizer + doc matrix |
encoder.json |
vocab + idf + sparse docs for the static Space |
Query the same way the Space does: vectorize the user text with this vocabulary, cosine vs documents, then write with Gemini on findexpert.ir (or future Qwen2.5-0.5B LoRA from ecoai-sft).
Dense upgrade later: encode passage: / query: with intfloat/multilingual-e5-small — do not replace this repo with forked MiniLM weights.