RooCode codesearch feature does not work with LM Studio

[christopherowen]

App Version

3.19.4

API Provider

LM Studio

Model Used

nomic-embed-code

🔁 Steps to Reproduce

On MacOS Sequoia 15.5, using LM Studio 0.3.16.

LM Studio will produce errors during this Indexing Process of:

Unexpected endpoint or method. (POST /embeddings). Returning 200 anyway

testing the API with python returns a result and no error in the LM Studio console:

# Make sure to `pip install openai` first
from openai import OpenAI
client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")

def get_embedding(text, model="nomic-embed-code"):
   return client.embeddings.create(input = [text], model=model).data[0].embedding

print(get_embedding("Once upon a time, there was a cat."))

Also, testing with curl produces a result and no error in the LM Studio console:

curl http://127.0.0.1:1234/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "model": "text-embedding-nomic-embed-code",
    "input": "Some text to embed"
  }'

Steps to reproduce:

• Download nomic-embed-code to LM Studio: https://huggingface.co/lmstudio-community/nomic-embed-code-GGUF

• Edit the model settings and set 'Override Domain Type' to 'Text Embedding'

• Launch Qdrant docker using the following docker-compose.yaml:

services:
  qdrant:
    image: qdrant/qdrant:latest
    restart: always
    container_name: qdrant
    ports:
      - 6333:6333
      - 6334:6334
    expose:
      - 6333
      - 6334
      - 6335
    configs:
      - source: qdrant_config
        target: /qdrant/config/production.yaml
    volumes:
      - ./qdrant_data:/qdrant/storage

configs:
  qdrant_config:
    content: |
      log_level: INFO

• Visit RooCode Experimental features

• Set Enable Codebase Indexing

• Embeddings Provider: OpenAI Compatible

• Base URL: http://localhost:1234 <- default LM Studio

• API Key: fakekey

• Model: nomic-embed-code

• Embedding Dimension: tried both blank and 3584

• Qdrant URL: http://localhost:6333

• Qdrant Key: provided during Qdrant first launch

• click Save

• click Start Indexing

Some debugging that was attempted:

• As I do not know what Dimensions are, I tried setting it to the qwen2.embedding_length of 3584.
• I tried setting the context window of the model to 3584.

💥 Outcome Summary

I expected the indexing operation to produce an index of my codebase. Instead there were errors during the indexing process in LM Studio and an empty index was produced in Qdrant with 0 Points, 8 Segments, and 1 Shard.

📄 Relevant Logs or Errors (Optional)

LM Studio Developer Logs:

2025-06-07 02:36:18 [DEBUG] 
Received request: POST to /embeddings with body  {
  "input": [
    "public function down()\n    {\n\n        $container =... <Truncated in logs> ...\n        $container['db']->query($query);\n\n\n\n    }",

( more lines of input... )

  ],
  "model": "nomic-embed-code",
  "encoding_format": "base64"
}
2025-06-07 02:36:18 [ERROR] 
Unexpected endpoint or method. (POST /embeddings). Returning 200 anyway

[christopherowen]

[christopherowen]

This is related to #4066

[christopherowen]

Indexing via Ollama was working with nomic-embed-text earlier.

[dosubot]

Hi @christopherowen! I'm Dosu and I’m helping the Roo-Code team.

Thanks for clarifying that indexing worked with nomic-embed-text via Ollama. Based on the docs, nomic-embed-text is the recommended embedding model and is known to work with RooCode and Ollama for codebase indexing. There’s no specific documentation or troubleshooting info for nomic-embed-code with LM Studio, and the error you’re seeing (“Unexpected endpoint or method. (POST /embeddings). Returning 200 anyway”) isn’t covered in the RooCode docs.

Given that your API tests with Python and curl succeed, but RooCode indexing fails and Qdrant remains empty, it’s possible there’s a subtle incompatibility or configuration mismatch between RooCode’s OpenAI-compatible embedder and how LM Studio exposes nomic-embed-code. If you haven’t already, double-check the model name and embedding dimension, and try using nomic-embed-text with LM Studio if possible, since that model is explicitly recommended and tested for embedding workflows in RooCode docs.

If you need to use nomic-embed-code specifically, you might want to compare the API request payloads between RooCode and your working curl/Python examples to see if there’s a difference in model naming or input formatting. If you spot anything, let me know!

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[christopherowen]

The model name was checked via the test python script and curl calls.

It is not clear what Embedding Dimension should be set to.

Using nomic-embed-text from https://huggingface.co/nomic-ai/nomic-embed-text-v1.5-GGUF with the above setup via LM Studio will produce the same failure.

I cannot find differences in the test python script and the API calls made by RooCode apart from the content in the input section of the json.

[adamhill]

Roo is using the OAI SDK. The cURL's work and are pulled directly from the API docs..... Grrr.

The only other thing I can think of is that the test string is way below a "normal" chunking size (1000 or so). So we get a false success from the simple test. But Roo is using some size it doesnt like when sending.

[christopherowen]

perhaps some tunables with which we can limit the length of inputs or number of inputs in the embedding calls could help confirm this?

[christopherowen]

I was able to resolve this by:

1. Changing the Base URL to http://localhost:1234/v1 <-- the /v1 was missing
2. Setting the dimensions to a small number and reading the error message in the qdrant console:

[qdrant] | 2025-06-07T07:39:06.864080Z  WARN collection::collection_manager::collection_updater: Update operation declined: Wrong input: Vector dimension error: expected dim: 512, got 896  

for nomic-embed-code, the Dimension should be set to 896.