Merge branch 'main' into fix/lmstudio-embeddings-base-url

The embedder fix in this branch left the sibling LLM still hardcoded, so
a self-hosted LM Studio behind a non-default host worked for embeddings
and silently fell back to localhost for generation.

LMStudioConfig baked "http://localhost:1234/v1" into __init__, which
makes any env-var lookup downstream dead code. The field now defaults to
None and LMStudioLLM resolves config value, then LMSTUDIO_BASE_URL, then
the localhost default, matching the embedder and DeepSeek.

mem0-ts/src/oss/src/embeddings/lmstudio.ts

@@ -12,7 +12,11 @@ export class LMStudioEmbedder implements Embedder {
   private model: string;
 
   constructor(config: EmbeddingConfig) {
-    const baseURL = config.baseURL ?? config.url ?? DEFAULT_BASE_URL;
+    const baseURL =
+      config.baseURL ??
+      config.url ??
+      process.env.LMSTUDIO_BASE_URL ??
+      DEFAULT_BASE_URL;
     const apiKey = config.apiKey || DEFAULT_LMSTUDIO_API_KEY;
     this.openai = new OpenAI({ apiKey, baseURL: String(baseURL) });
     this.model = config.model || DEFAULT_MODEL;

mem0-ts/src/oss/src/llms/lmstudio.ts

@@ -12,7 +12,8 @@ export class LMStudioLLM extends OpenAILLM {
     super({
       ...config,
       apiKey: config.apiKey || DEFAULT_LMSTUDIO_API_KEY,
-      baseURL: config.baseURL ?? DEFAULT_BASE_URL,
+      baseURL:
+        config.baseURL ?? process.env.LMSTUDIO_BASE_URL ?? DEFAULT_BASE_URL,
       model: config.model || DEFAULT_MODEL,
     });
   }

mem0-ts/src/oss/tests/lmstudio-embedder.test.ts

@@ -10,14 +10,64 @@ const mockCreate = jest.fn().mockResolvedValue({
   data: [{ embedding: mockEmbedding }],
 });
 
+let capturedConstructorArgs: any;
+
 jest.mock("openai", () => {
-  return jest.fn().mockImplementation(() => ({
-    embeddings: { create: mockCreate },
-  }));
+  return jest.fn().mockImplementation((args: any) => {
+    capturedConstructorArgs = args;
+    return {
+      embeddings: { create: mockCreate },
+    };
+  });
 });
 
 describe("LMStudioEmbedder (unit)", () => {
-  beforeEach(() => mockCreate.mockClear());
+  beforeEach(() => {
+    mockCreate.mockClear();
+    capturedConstructorArgs = undefined;
+    delete process.env.LMSTUDIO_BASE_URL;
+  });
+
+  afterEach(() => {
+    delete process.env.LMSTUDIO_BASE_URL;
+  });
+
+  it("uses the default base URL when neither config nor env var is set", async () => {
+    const embedder = new LMStudioEmbedder({ model: "test-model" });
+
+    await embedder.embed("hello");
+
+    expect(capturedConstructorArgs).toMatchObject({
+      baseURL: "http://localhost:1234/v1",
+    });
+  });
+
+  it("uses LMSTUDIO_BASE_URL when config does not provide a base URL", async () => {
+    process.env.LMSTUDIO_BASE_URL = "http://example.test:5678/v1";
+
+    const embedder = new LMStudioEmbedder({ model: "test-model" });
+
+    await embedder.embed("hello");
+
+    expect(capturedConstructorArgs).toMatchObject({
+      baseURL: "http://example.test:5678/v1",
+    });
+  });
+
+  it("prefers an explicit config baseURL over LMSTUDIO_BASE_URL", async () => {
+    process.env.LMSTUDIO_BASE_URL = "http://example.test:5678/v1";
+
+    const embedder = new LMStudioEmbedder({
+      model: "test-model",
+      baseURL: "http://config.test:9999/v1",
+    });
+
+    await embedder.embed("hello");
+
+    expect(capturedConstructorArgs).toMatchObject({
+      baseURL: "http://config.test:9999/v1",
+    });
+  });
 
   it("embed() calls OpenAI with encoding_format float and returns vector", async () => {
     const embedder = new LMStudioEmbedder({

mem0-ts/src/oss/tests/lmstudio-llm.test.ts

@@ -5,16 +5,56 @@
 
 import { LMStudioLLM } from "../src/llms/lmstudio";
 
+let capturedConstructorArgs: any;
 const mockCreate = jest.fn();
 
 jest.mock("openai", () => {
-  return jest.fn().mockImplementation(() => ({
-    chat: { completions: { create: mockCreate } },
-  }));
+  return jest.fn().mockImplementation((args: any) => {
+    capturedConstructorArgs = args;
+    return {
+      chat: { completions: { create: mockCreate } },
+    };
+  });
 });
 
 describe("LMStudioLLM (unit)", () => {
-  beforeEach(() => mockCreate.mockClear());
+  beforeEach(() => {
+    mockCreate.mockClear();
+    capturedConstructorArgs = undefined;
+    delete process.env.LMSTUDIO_BASE_URL;
+  });
+
+  afterEach(() => {
+    delete process.env.LMSTUDIO_BASE_URL;
+  });
+
+  it("uses the default base URL when neither config nor env var is set", () => {
+    new LMStudioLLM({});
+
+    expect(capturedConstructorArgs).toMatchObject({
+      baseURL: "http://localhost:1234/v1",
+    });
+  });
+
+  it("uses LMSTUDIO_BASE_URL when config does not provide a base URL", () => {
+    process.env.LMSTUDIO_BASE_URL = "http://example.test:5678/v1";
+
+    new LMStudioLLM({});
+
+    expect(capturedConstructorArgs).toMatchObject({
+      baseURL: "http://example.test:5678/v1",
+    });
+  });
+
+  it("prefers an explicit config baseURL over LMSTUDIO_BASE_URL", () => {
+    process.env.LMSTUDIO_BASE_URL = "http://example.test:5678/v1";
+
+    new LMStudioLLM({ baseURL: "http://config.test:9999/v1" });
+
+    expect(capturedConstructorArgs).toMatchObject({
+      baseURL: "http://config.test:9999/v1",
+    });
+  });
 
   it("generateResponse() returns a text response", async () => {
     mockCreate.mockResolvedValueOnce({

mem0/configs/embeddings/base.py

@@ -34,7 +34,7 @@ class BaseEmbedderConfig(ABC):
         # Gemini specific
         output_dimensionality: Optional[str] = None,
         # LM Studio specific
-        lmstudio_base_url: Optional[str] = "http://localhost:1234/v1",
+        lmstudio_base_url: Optional[str] = None,
         # AWS Bedrock specific
         aws_access_key_id: Optional[str] = None,
         aws_secret_access_key: Optional[str] = None,
@@ -70,7 +70,8 @@ class BaseEmbedderConfig(ABC):
         :type memory_update_embedding_type: Optional[str], optional
         :param memory_search_embedding_type: The type of embedding to use for the search memory action, defaults to None
         :type memory_search_embedding_type: Optional[str], optional
-        :param lmstudio_base_url: LM Studio base URL to be use, defaults to "http://localhost:1234/v1"
+        :param lmstudio_base_url: LM Studio base URL to be use, defaults to None
+            (resolved by the provider from LMSTUDIO_BASE_URL, then "http://localhost:1234/v1")
         :type lmstudio_base_url: Optional[str], optional
         """
 

mem0/configs/llms/lmstudio.py

@@ -39,6 +39,7 @@ class LMStudioConfig(BaseLlmConfig):
             vision_details: Vision detail level, defaults to "auto"
             http_client_proxies: HTTP client proxy settings, defaults to None
             lmstudio_base_url: LM Studio base URL, defaults to None
+                (resolved by the provider from LMSTUDIO_BASE_URL, then "http://localhost:1234/v1")
             lmstudio_response_format: LM Studio response format, defaults to None
         """
         # Initialize base parameters
@@ -55,5 +56,5 @@ class LMStudioConfig(BaseLlmConfig):
         )
 
         # LM Studio-specific parameters
-        self.lmstudio_base_url = lmstudio_base_url or "http://localhost:1234/v1"
+        self.lmstudio_base_url = lmstudio_base_url
         self.lmstudio_response_format = lmstudio_response_format

mem0/embeddings/lmstudio.py

@@ -1,3 +1,4 @@
+import os
 from typing import Literal, Optional
 
 from openai import OpenAI
@@ -13,6 +14,9 @@ class LMStudioEmbedding(EmbeddingBase):
         self.config.model = self.config.model or "nomic-ai/nomic-embed-text-v1.5-GGUF/nomic-embed-text-v1.5.f16.gguf"
         self.config.embedding_dims = self.config.embedding_dims or 1536
         self.config.api_key = self.config.api_key or "lm-studio"
+        self.config.lmstudio_base_url = (
+            self.config.lmstudio_base_url or os.getenv("LMSTUDIO_BASE_URL") or "http://localhost:1234/v1"
+        )
 
         self.client = OpenAI(base_url=self.config.lmstudio_base_url, api_key=self.config.api_key)
 

mem0/embeddings/vertexai.py

@@ -8,14 +8,6 @@ from mem0.embeddings.base import EmbeddingBase
 from mem0.utils.gcp_auth import GCPAuthenticator
 
 
-def _max_instances_per_request(model: str) -> int:
-    # Vertex AI's per-request input cap depends on the model family: gemini-embedding-*
-    # accepts exactly one input text per predict() call, while text-embedding-004/005 and
-    # text-multilingual-embedding-002 accept up to 250.
-    # https://docs.cloud.google.com/vertex-ai/docs/samples/aiplatform-sdk-embedding
-    return 1 if model.startswith("gemini-embedding") else 250
-
-
 class VertexAIEmbedding(EmbeddingBase):
     def __init__(self, config: Optional[BaseEmbedderConfig] = None):
         super().__init__(config)
@@ -80,9 +72,8 @@ class VertexAIEmbedding(EmbeddingBase):
                 raise ValueError(f"Invalid memory action: {memory_action}")
             embedding_type = self.embedding_types[memory_action]
         all_embeddings = []
-        chunk_size = _max_instances_per_request(self.config.model)
-        for i in range(0, len(texts), chunk_size):
-            chunk = texts[i : i + chunk_size]
+        for i in range(0, len(texts), 250):
+            chunk = texts[i : i + 250]
             inputs = [TextEmbeddingInput(text=t, task_type=embedding_type) for t in chunk]
             results = self.model.get_embeddings(texts=inputs, output_dimensionality=self.config.embedding_dims)
             all_embeddings.extend(r.values for r in results)

mem0/llms/lmstudio.py

@@ -1,4 +1,5 @@
 import json
+import os
 from typing import Dict, List, Optional, Union
 
 from openai import OpenAI
@@ -37,6 +38,9 @@ class LMStudioLLM(LLMBase):
             or "lmstudio-community/Meta-Llama-3.1-70B-Instruct-GGUF/Meta-Llama-3.1-70B-Instruct-IQ2_M.gguf"
         )
         self.config.api_key = self.config.api_key or "lm-studio"
+        self.config.lmstudio_base_url = (
+            self.config.lmstudio_base_url or os.getenv("LMSTUDIO_BASE_URL") or "http://localhost:1234/v1"
+        )
 
         self.client = OpenAI(base_url=self.config.lmstudio_base_url, api_key=self.config.api_key)
 

tests/embeddings/test_lm_studio_embeddings.py

@@ -6,7 +6,6 @@ from mem0.configs.embeddings.base import BaseEmbedderConfig
 from mem0.embeddings.lmstudio import LMStudioEmbedding
 
 
-
 @pytest.fixture
 def mock_lm_studio_client():
     with patch("mem0.embeddings.lmstudio.OpenAI") as mock_openai:
@@ -80,3 +79,31 @@ def test_embed_batch_count_mismatch_raises(mock_lm_studio_client):
 
     with pytest.raises(ValueError, match="returned 1 embeddings for 2 texts"):
         embedder.embed_batch(["first text", "second text"])
+
+
+def test_base_url_honors_lmstudio_base_url_env(mock_lm_studio_client, monkeypatch):
+    monkeypatch.setenv("LMSTUDIO_BASE_URL", "http://lmstudio.internal:9000/v1")
+
+    with patch("mem0.embeddings.lmstudio.OpenAI") as mock_openai:
+        LMStudioEmbedding(BaseEmbedderConfig())
+
+    assert mock_openai.call_args.kwargs["base_url"] == "http://lmstudio.internal:9000/v1"
+
+
+def test_base_url_prefers_explicit_config_over_env(mock_lm_studio_client, monkeypatch):
+    monkeypatch.setenv("LMSTUDIO_BASE_URL", "http://lmstudio.internal:9000/v1")
+
+    with patch("mem0.embeddings.lmstudio.OpenAI") as mock_openai:
+        LMStudioEmbedding(BaseEmbedderConfig(lmstudio_base_url="http://explicit:1234/v1"))
+
+    assert mock_openai.call_args.kwargs["base_url"] == "http://explicit:1234/v1"
+
+
+def test_base_url_falls_back_to_localhost_default(mock_lm_studio_client, monkeypatch):
+    monkeypatch.delenv("LMSTUDIO_BASE_URL", raising=False)
+
+    with patch("mem0.embeddings.lmstudio.OpenAI") as mock_openai:
+        embedder = LMStudioEmbedding(BaseEmbedderConfig())
+
+    assert mock_openai.call_args.kwargs["base_url"] == "http://localhost:1234/v1"
+    assert embedder.config.lmstudio_base_url == "http://localhost:1234/v1"

tests/embeddings/test_vertexai_embeddings.py

@@ -163,9 +163,7 @@ def test_invalid_memory_action(mock_text_embedding_model, mock_config):
 
 @patch("mem0.embeddings.vertexai.TextEmbeddingModel")
 def test_embed_batch_single_call(mock_text_embedding_model, mock_os_environ, mock_config):
-    # text-embedding-005 accepts up to 250 inputs per request, so 2 texts fit in one call.
-    # (gemini-embedding-* only accepts 1 input per request; see test_embed_batch_default_model_sends_one_text_per_call.)
-    mock_config.return_value.model = "text-embedding-005"
+    mock_config.return_value.model = "gemini-embedding-001"
     mock_config.return_value.embedding_dims = 256
 
     config = mock_config()
@@ -200,9 +198,7 @@ def test_embed_batch_empty_list(mock_text_embedding_model, mock_os_environ, mock
 
 @patch("mem0.embeddings.vertexai.TextEmbeddingModel")
 def test_embed_batch_count_mismatch_raises(mock_text_embedding_model, mock_os_environ, mock_config):
-    # Use a 250-cap model so both texts land in a single chunk/call, keeping the mismatch
-    # (1 embedding returned for that one call) triggered by that single call.
-    mock_config.return_value.model = "text-embedding-005"
+    mock_config.return_value.model = "gemini-embedding-001"
     mock_config.return_value.embedding_dims = 256
 
     config = mock_config()
@@ -265,8 +261,8 @@ def test_embed_batch_invalid_memory_action_raises(mock_text_embedding_model, moc
 
 @patch("mem0.embeddings.vertexai.TextEmbeddingModel")
 def test_embed_batch_chunking_triggers_two_api_calls(mock_text_embedding_model, mock_os_environ, mock_config):
-    """text-embedding-005 accepts up to 250 inputs/request: 300 texts -> 2 calls (chunks of 250 and 50)."""
-    mock_config.return_value.model = "text-embedding-005"
+    """300 texts must produce exactly 2 get_embeddings calls (chunks of 250 and 50)."""
+    mock_config.return_value.model = "gemini-embedding-001"
     mock_config.return_value.embedding_dims = 256
 
     config = mock_config()
@@ -282,25 +278,3 @@ def test_embed_batch_chunking_triggers_two_api_calls(mock_text_embedding_model,
 
     assert mock_text_embedding_model.from_pretrained.return_value.get_embeddings.call_count == 2
     assert len(result) == 300
-
-
-@patch("mem0.embeddings.vertexai.TextEmbeddingModel")
-def test_embed_batch_default_model_sends_one_text_per_call(mock_text_embedding_model, mock_os_environ, mock_config):
-    """gemini-embedding-001 (default) accepts 1 input/request: 300 texts -> 300 calls of 1 text each."""
-    mock_config.return_value.model = "gemini-embedding-001"
-    mock_config.return_value.embedding_dims = 256
-
-    config = mock_config()
-    embedder = VertexAIEmbedding(config)
-
-    def make_single_response(texts, output_dimensionality):
-        assert len(texts) == 1
-        return [Mock(values=[0.1, 0.2]) for _ in texts]
-
-    mock_text_embedding_model.from_pretrained.return_value.get_embeddings.side_effect = make_single_response
-
-    texts = [f"text {i}" for i in range(300)]
-    result = embedder.embed_batch(texts)
-
-    assert mock_text_embedding_model.from_pretrained.return_value.get_embeddings.call_count == 300
-    assert len(result) == 300

tests/llms/test_lm_studio.py

@@ -2,6 +2,7 @@ from unittest.mock import Mock, patch
 
 import pytest
 
+from mem0.configs.llms.base import BaseLlmConfig
 from mem0.configs.llms.lmstudio import LMStudioConfig
 from mem0.llms.lmstudio import LMStudioLLM
 
@@ -70,3 +71,40 @@ def test_generate_response_specifying_response_format(mock_lm_studio_client):
     )
 
     assert response == "I'm doing well, thank you for asking!"
+
+
+def test_base_url_honors_lmstudio_base_url_env(monkeypatch):
+    monkeypatch.setenv("LMSTUDIO_BASE_URL", "http://lmstudio.internal:9000/v1")
+
+    with patch("mem0.llms.lmstudio.OpenAI") as mock_openai:
+        LMStudioLLM(LMStudioConfig())
+
+    assert mock_openai.call_args.kwargs["base_url"] == "http://lmstudio.internal:9000/v1"
+
+
+def test_base_url_prefers_explicit_config_over_env(monkeypatch):
+    monkeypatch.setenv("LMSTUDIO_BASE_URL", "http://lmstudio.internal:9000/v1")
+
+    with patch("mem0.llms.lmstudio.OpenAI") as mock_openai:
+        LMStudioLLM(LMStudioConfig(lmstudio_base_url="http://explicit:1234/v1"))
+
+    assert mock_openai.call_args.kwargs["base_url"] == "http://explicit:1234/v1"
+
+
+def test_base_url_falls_back_to_localhost_default(monkeypatch):
+    monkeypatch.delenv("LMSTUDIO_BASE_URL", raising=False)
+
+    with patch("mem0.llms.lmstudio.OpenAI") as mock_openai:
+        llm = LMStudioLLM(LMStudioConfig())
+
+    assert mock_openai.call_args.kwargs["base_url"] == "http://localhost:1234/v1"
+    assert llm.config.lmstudio_base_url == "http://localhost:1234/v1"
+
+
+def test_base_url_honors_env_via_base_llm_config(monkeypatch):
+    monkeypatch.setenv("LMSTUDIO_BASE_URL", "http://lmstudio.internal:9000/v1")
+
+    with patch("mem0.llms.lmstudio.OpenAI") as mock_openai:
+        LMStudioLLM(BaseLlmConfig(model="some-model"))
+
+    assert mock_openai.call_args.kwargs["base_url"] == "http://lmstudio.internal:9000/v1"