Fixup bert python bindings.

gpt4all-backend/bert.cpp

@@ -14,6 +14,7 @@
 #include <regex>
 #include <thread>
 #include <algorithm>
+#include <numeric>
 
 //#define DEBUG_BERT
 
@@ -462,11 +463,6 @@ void bert_eval(
     ggml_set_f32(sum, 1.0f / N);
     inpL = ggml_mul_mat(ctx0, inpL, sum);
 
-    // normalizer
-    ggml_tensor *length = ggml_sqrt(ctx0,
-                                    ggml_sum(ctx0, ggml_sqr(ctx0, inpL)));
-    inpL = ggml_scale(ctx0, inpL, ggml_div(ctx0, ggml_new_f32(ctx0, 1.0f), length));
-
     ggml_tensor *output = inpL;
     // run the computation
     ggml_build_forward_expand(&gf, output);
@@ -987,6 +983,9 @@ std::vector<float> Bert::embedding(const std::string &text)
     }
 
     std::transform(embeddingsSum.begin(), embeddingsSum.end(), embeddingsSum.begin(), [embeddingsSumTotal](float num){ return num / embeddingsSumTotal; });
+    double magnitude = std::sqrt(std::inner_product(embeddingsSum.begin(), embeddingsSum.end(), embeddingsSum.begin(), 0.0));
+    for (auto &value : embeddingsSum)
+        value /= magnitude;
     std::vector<float> finalEmbeddings(embeddingsSum.begin(), embeddingsSum.end());
     return finalEmbeddings;
 }

gpt4all-bindings/python/gpt4all/__init__.py

@@ -1,2 +1,2 @@
-from .gpt4all import GPT4All, embed  # noqa
+from .gpt4all import GPT4All, Embedder  # noqa
 from .pyllmodel import LLModel  # noqa

gpt4all-bindings/python/gpt4all/gpt4all.py

@@ -15,20 +15,26 @@ from . import pyllmodel
 # TODO: move to config
 DEFAULT_MODEL_DIRECTORY = os.path.join(str(Path.home()), ".cache", "gpt4all").replace("\\", "\\\\")
 
-def embed(
-    text: str
-) -> list[float]:
-    """
-    Generate an embedding for all GPT4All.
+class Embedder:
+    def __init__(
+        self
+    ):
+        self.gpt4all = GPT4All(model_name='ggml-all-MiniLM-L6-v2-f16.bin', n_threads=8)
+
+    def embed(
+        self,
+        text: str
+    ) -> list[float]:
+        """
+        Generate an embedding for all GPT4All.
 
-    Args:
-        text: The text document to generate an embedding for.
+        Args:
+            text: The text document to generate an embedding for.
 
-    Returns:
-        An embedding of your document of text.
-    """
-    model = GPT4All(model_name='ggml-all-MiniLM-L6-v2-f16.bin')
-    return model.model.generate_embedding(text)
+        Returns:
+            An embedding of your document of text.
+        """
+        return self.gpt4all.model.generate_embedding(text)
 
 class GPT4All:
     """

gpt4all-bindings/python/gpt4all/pyllmodel.py

@@ -253,7 +253,7 @@ class LLModel:
         embedding_size = ctypes.c_size_t()
         c_text = ctypes.c_char_p(text.encode('utf-8'))
         embedding_ptr = llmodel.llmodel_embedding(self.model, c_text, ctypes.byref(embedding_size))
-        embedding_array = ctypes.cast(embedding_ptr, ctypes.POINTER(ctypes.c_float * embedding_size.value)).contents
+        embedding_array = [embedding_ptr[i] for i in range(embedding_size.value)]
         llmodel.llmodel_free_embedding(embedding_ptr)
         return list(embedding_array)