PWRWork local semantic search model
====================================

Files
  pwr-minilm-static.bin  7,935,736 bytes  static token-embedding table (30,522 rows x 256 int8 + one float scale per row)
  vocab.txt                231,508 bytes  WordPiece vocabulary (uncased, 30,522 tokens)

What it is
  A static table distilled from the sentence-transformers model all-MiniLM-L6-v2: every vocabulary
  token was run through the original model alone ([CLS] token [SEP], mean pooled), the 384-dimension
  vectors were reduced to 256 dimensions with PCA (95.9% of the variance kept) and stored as int8.
  pwr_index_search (mode semantic or hybrid) splits text into words, applies WordPiece with this vocabulary,
  averages the token rows and compares the normalised vectors by cosine similarity. Everything runs
  inside the Editor process on the CPU: no network, no API key, no Python, no GPU.
  The distillation script is Tools/distill.py, its check against the full model Tools/evaluate.py.

Origin and license
  all-MiniLM-L6-v2 (https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2), by the
  sentence-transformers project, fine-tuned from nreimers/MiniLM-L6-H384-uncased (Microsoft MiniLM),
  vocabulary from Google's BERT uncased. License: Apache License 2.0 (Licensing/Apache-2.0.txt).
  This table is a modified derivative: it contains no original network weights, only the per-token
  outputs described above. Notice in Licensing/ThirdPartyNotices.txt.

Distribution
  The two files ship inside the plugin package (Content/Semantic). If they are missing or damaged,
  pwr_index_search falls back to its ranked keyword search and says so (semantic.available=false).
