Document the GGUF alongside the ONNX

README.md

@@ -1,6 +1,17 @@
 A small English embedding model by BAAI (Beijing Academy of Artificial Intelligence).
 
-This repository hosts the ONNX version of the model used by Understand to generate embeddings for semantic search, allowing users to search their codebase by meaning rather than exact keyword matches.
+This repository hosts the versions of the model used by Understand to generate embeddings for semantic search, allowing users to search their codebase by meaning rather than exact keyword matches.
+
+- `bge-small-en-v1.5-f16.gguf` — GGUF (F16), served by ullama with `--embeddings`.
+  Used by Understand 2026 and later. 384 dimensions, CLS pooling, 512-token context.
+- `bge-small-en-v1.5.onnx` + `bge-small-en-v1.5-tokenizer.json` — ONNX, used by the
+  retired undaiserver. Kept for older releases.
+
+The GGUF was converted from `BAAI/bge-small-en-v1.5` with llama.cpp's
+`convert_hf_to_gguf.py`; its vectors match the reference implementation
+(cosine similarity 1.00000). Note the ONNX path used mean pooling while the
+GGUF uses BGE's specified CLS pooling, so indexes built with one are not
+comparable with the other.
 
 ---
 license: mit