Removed the dependency on numpy.

llmclient/embeddings.py

@@ -1,10 +1,11 @@
 import asyncio
 from abc import ABC, abstractmethod
+from collections import Counter
 from enum import StrEnum
+from itertools import chain
 from typing import Any
 
 import litellm
-import numpy as np
 import tiktoken
 from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
 
@@ -171,13 +172,9 @@ async def embed_documents(self, texts: list[str]) -> list[list[float]]:
         enc_batch = self.enc.encode_ordinary_batch(texts)
         # now get frequency of each token rel to length
         return [
-            (
-                np.bincount([xi % self.ndim for xi in x], minlength=self.ndim).astype(
-                    float
-                )
-                / len(x)
-            ).tolist()
+            [token_counts.get(xi, 0) / len(x) for xi in range(self.ndim)]
             for x in enc_batch
+            if (token_counts := Counter(xi % self.ndim for xi in x))
         ]
 
 
@@ -199,7 +196,11 @@ async def embed_documents(self, texts: list[str]) -> list[list[float]]:
         all_embeds = await asyncio.gather(
             *[m.embed_documents(texts) for m in self.models]
         )
-        return np.concatenate(all_embeds, axis=1).tolist()
+
+        return [
+            list(chain.from_iterable(embed_group))
+            for embed_group in zip(*all_embeds, strict=True)
+        ]
 
     def set_mode(self, mode: EmbeddingModes) -> None:
         # Set mode for all component models
@@ -217,6 +218,7 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
         try:
+            import numpy as np  # noqa: F401
             from sentence_transformers import SentenceTransformer
         except ImportError as exc:
             raise ImportError(
@@ -240,6 +242,8 @@ async def embed_documents(self, texts: list[str]) -> list[list[float]]:
         Returns:
             A list of embedding vectors.
         """
+        import numpy as np
+
         # Extract additional configurations if needed
         batch_size = self.config.get("batch_size", 32)
         device = self.config.get("device", "cpu")

pyproject.toml

@@ -26,7 +26,6 @@ dependencies = [
     "fhaviary>=0.8.2",  # For core namespace
     "limits",
     "litellm>=1.44",  # For LITELLM_LOG addition
-    "numpy",
     "pydantic~=2.0,>=2.10.1,<2.10.2",
     "tiktoken>=0.4.0",
     "typing-extensions; python_version <= '3.11'",  # for typing.override
@@ -40,7 +39,7 @@ requires-python = ">=3.11"
 
 [project.optional-dependencies]
 dev = [
-    "fh-llm-client[image,local]",
+    "fh-llm-client[local]",
     "fhaviary[xml]",
     "ipython>=8",  # Pin to keep recent
     "mypy>=1.8",  # Pin for mutable-override
@@ -58,10 +57,8 @@ dev = [
     "python-dotenv",
     "refurb>=2",  # Pin to keep recent
 ]
-image = [
-    "Pillow",
-]
 local = [
+    "numpy",
     "sentence-transformers",
 ]