OpenAIEmbeddings does not allow setting encoding_format, causing incompatibility with LM Studio (returns float[], not base64)

[GregorBiswanger]

Checked other resources

• I added a very descriptive title to this issue.
• I searched the LangChain.js documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain.js rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).

Example Code

WORKS with LM Studio if encoding_format is set manually (value doesn't matter)

const openAiClient = new OpenAIClient(createLMStudioConfig());

async function getEmbeddings(text) {
  const response = await openAiClient.embeddings.create({
    model: 'none',
    input: text,
    encoding_format: 'base64',
  });

  return response.data[0].embedding;
}

DOES NOT WORK with OpenAIEmbeddings – no way to set encoding_format

  const embeddings = new OpenAIEmbeddings({ ...createLMStudioConfig(), model: '' });

  const vectorStore = await QdrantVectorStore.fromExistingCollection(embeddings, {
    url: 'http://localhost:6333',
    collectionName: 'eu-ai',
  });

Error Message and Stack Trace (if applicable)

[Running] node "c:\projects\rag-sample\index.mjs"
✅ Server running at http://localhost:11434
file:///c:/projects/rag-sample/node_modules/@qdrant/openapi-typescript-fetch/dist/esm/fetcher.js:169
throw new fun.Error(err);
^

ApiError: Bad Request
at Object.fun [as searchPoints] (file:///c:/projects/rag-sample/node_modules/@qdrant/openapi-typescript-fetch/dist/esm/fetcher.js:169:23)
at process.processTicksAndRejections (node:internal/process/task_queues:105:5)
at async QdrantClient.search (file:///c:/projects/rag-sample/node_modules/@qdrant/js-client-rest/dist/esm/qdrant-client.js:165:26)
at async QdrantVectorStore.similaritySearchVectorWithScore (file:///c:/projects/rag-sample/node_modules/@langchain/qdrant/dist/vectorstores.js:165:25)
at async QdrantVectorStore.similaritySearch (file:///c:/projects/rag-sample/node_modules/@langchain/core/dist/vectorstores.js:260:25)
at async file:///c:/projects/rag-sample/index.mjs:55:29 {
headers: Headers {},
url: 'http://localhost:6333/collections/eu-ai/points/search',
status: 400,
statusText: 'Bad Request',
data: {
status: {
error: 'Wrong input: Vector dimension error: expected dim: 768, got 192'
},
time: 0.001095872
}
}

Node.js v22.14.0

[Done] exited with code=1 in 6.972 seconds

Description

I'm using OpenAIEmbeddings from LangChainJS in combination with LM Studio as the embedding backend. LM Studio provides a fully OpenAI-compatible API, but it always returns raw float[] embeddings, regardless of the encoding_format parameter.

Here’s the problem:

• The LangChainJS OpenAIEmbeddings class does not expose any option to set encoding_format.
• The underlying openai SDK (v4.x) automatically defaults to encoding_format: "base64" when this is not set.
• As a result, LangChain expects base64-encoded embeddings, but LM Studio returns float[], leading to exceptions.

The only way to make this work is to explicitly set encoding_format in the SDK call and interestingly, it works no matter what value you set ("float" or "base64"), because LM Studio ignores the parameter anyway and always returns float arrays.

However, since OpenAIEmbeddings does not allow setting encoding_format, I had to extend the class with a custom wrapper just to inject this parameter.

System Info

@langchain/community: 0.3.43
@langchain/core: 0.3.56
@langchain/openai: 0.5.10
@langchain/qdrant: 0.1.2
Node: v22.14.0
Platform: Windows 11
Embedding backend: LM-Studio (localhost)

[erbg]

same here.