add max size

libs/client_infinity/infinity_client/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "infinity_client"
-version = "0.0.71"
+version = "0.0.72"
 description = "A client library for accessing ♾️ Infinity - Embedding Inference Server"
 authors = []
 readme = "README.md"

libs/infinity_emb/infinity_emb/env.py

@@ -189,7 +189,15 @@ def cache_dir(self) -> Path:
 
     @cached_property
     def queue_size(self) -> int:
-        return int(self._optional_infinity_var("queue_size", default="32000"))
+        size = int(self._optional_infinity_var("queue_size", default="32000"))
+        assert size > 0, "INFINITY_QUEUE_SIZE must be a positive number"
+        return size
+
+    @cached_property
+    def max_client_batch_size(self) -> int:
+        size = int(self._optional_infinity_var("max_client_batch_size", default="2048"))
+        assert size > 0, "INFINITY_MAX_CLIENT_BATCH_SIZE must be a positive number"
+        return size
 
     @cached_property
     def permissive_cors(self):

libs/infinity_emb/infinity_emb/fastapi_schemas/pydantic_v2.py

@@ -1,4 +1,5 @@
 from pydantic import AnyUrl, HttpUrl, StringConstraints
+from infinity_emb.env import MANAGER
 
 __all__ = [
     "INPUT_STRING",
@@ -14,9 +15,9 @@
 INPUT_STRING = StringConstraints(max_length=8192 * 15, strip_whitespace=True)
 ITEMS_LIMIT = {
     "min_length": 1,
-    "max_length": 2048,
+    "max_length": MANAGER.max_client_batch_size,
 }
 ITEMS_LIMIT_SMALL = {
     "min_length": 1,
-    "max_length": 32,
+    "max_length": min(32, MANAGER.max_client_batch_size),
 }

libs/infinity_emb/pyproject.toml

@@ -1,7 +1,7 @@
 
 [tool.poetry]
 name = "infinity_emb"
-version = "0.0.71"
+version = "0.0.72"
 description = "Infinity is a high-throughput, low-latency REST API for serving text-embeddings, reranking models and clip."
 authors = ["michaelfeil <[email protected]>"]
 license = "MIT"