Bug: OpenAIEmbedder ignores baseURL config parameter

[zhezzma]

Bug Report: OpenAIEmbedder ignores baseURL config

Summary

The OpenAIEmbedder class in mem0 TypeScript SDK does not use the baseURL configuration parameter when initializing the OpenAI client, forcing users to manually set process.env.OPENAI_BASE_URL as a workaround.

Steps to Reproduce

1. Configure mem0 with a custom OpenAI-compatible API endpoint (e.g., NVIDIA API, local Ollama):

   const memory = new Memory({
     embedder: {
       provider: "openai",
       config: {
         apiKey: "sk-xxx",
         baseURL: "https://integrate.api.nvidia.com/v1",
         model: "nvidia/llama-3.1-nv-embedqa-1b-v1"
       }
     }
   });

2. Initialize the memory instance

3. Attempt to add or search memories

4. Error occurs because the OpenAI client connects to default OpenAI URL instead of the custom baseURL

Expected Behavior

The OpenAIEmbedder should respect the baseURL configuration parameter, similar to how OpenAILLM does.

Actual Behavior

The OpenAIEmbedder constructor only passes apiKey to the OpenAI client, ignoring baseURL:

var OpenAIEmbedder = class {
  constructor(config) {
    this.openai = new import_openai.default({ apiKey: config.apiKey }); // ❌ baseURL is ignored
    // ...
  }
}

Workaround

Users are forced to set environment variables manually before initializing mem0:

process.env.OPENAI_BASE_URL = customBaseURL;
process.env.OPENAI_API_BASE = customBaseURL;
const memory = new Memory({ /* config */ });

Comparison with OpenAILLM

The OpenAILLM class correctly implements baseURL support:

var OpenAILLM = class {
  constructor(config) {
    this.openai = new import_openai2.default({
      apiKey: config.apiKey,
      baseURL: config.baseURL  // ✅ Correctly uses baseURL
    });
    // ...
  }
}

Environment

• mem0ai version: 2.2.4
• Node.js version: 20.x
• OS: Linux

Proposed Fix

Update the OpenAIEmbedder class in src/oss/src/embeddings/openai.ts:

var OpenAIEmbedder = class {
  constructor(config) {
    this.openai = new import_openai.default({ 
      apiKey: config.apiKey,
      baseURL: config.baseURL  // ✅ Add this line
    });
    this.model = config.model || "text-embedding-3-small";
    this.embeddingDims = config.embeddingDims || 1536;
  }
  // ...
}

Impact

This affects all users who:

• Use custom OpenAI-compatible API endpoints (NVIDIA, local models, etc.)
• Cannot or prefer not to set global environment variables
• Want to keep configuration clean and explicit

Additional Notes

• The same issue may exist in AzureOpenAIEmbedder class
• Users should not need to rely on environment variable hacks for a standard configuration parameter
• This inconsistency between OpenAILLM and OpenAIEmbedder is confusing and error-prone

[superfat1988]

Thanks for reporting this — we can confirm the same behavior in a real OpenClaw + @mem0/openclaw-mem0 deployment (mem0ai 2.2.4, plugin 0.1.2).\n\n### What we observed\n- / / / were set to an OpenAI-compatible endpoint (SiliconFlow).\n- / still failed until we patched the runtime constructor to pass through base URL.\n- already respected , so this looked inconsistent specifically in embedder path.\n\n### Temporary workaround (validated)\nIn installed runtime files:\n- \n- \n\nwe changed embedder client init from:\n\nto:\n\n\nAfter restart, recall/capture path worked as expected in our environment.\n\n### Additional note\nIn our stack, once endpoint forwarding worked, we also had to ensure embedding dimensions matched vector-store collection size (1024 in our case) to avoid downstream query errors.\n\nHappy to test a fix build quickly if maintainers provide one. Appreciate the work on this project 🙏

[superfat1988]

Follow-up with exact fields/paths (previous comment had markdown escaping issues from shell):

• Config fields used in our OpenClaw plugin config:
  • oss.embedder.config.baseURL
  • oss.embedder.config.baseUrl
  • oss.embedder.config.openaiBaseUrl
  • oss.embedder.config.url
• Failing tools before hot patch:
  • memory_store
  • memory_search
• Inconsistent constructor path:
  • OpenAIEmbedder ignored base URL
  • OpenAILLM respected baseURL
• Runtime files patched:
  • node_modules/mem0ai/dist/oss/index.mjs
  • node_modules/mem0ai/dist/oss/index.js

Patch shape we validated:

new OpenAI({
  apiKey: config.apiKey,
  baseURL: config.baseURL || config.baseUrl || config.openaiBaseUrl || config.url || process.env.OPENAI_BASE_URL
})

After gateway restart, recall/capture worked in our OpenClaw environment.

If needed, we can provide a minimal reproducible config snippet and logs.

[kartik-mem0]

#4275

this issue has been resolved.