Upgrade `cocoindex[litellm]` dependency or filter `encoding_format=None` for OpenAI-compatible embeddings

[lemoay]

Summary

ccc index fails when provider: litellm is used with an OpenAI-compatible embeddings endpoint, because the outgoing embeddings request includes:

{
  "input": ["hello"],
  "model": "Qwen/Qwen3-Embedding-8B",
  "encoding_format": null
}

If encoding_format is omitted entirely, or changed to "float", the same request succeeds.

This appears to match LiteLLM's previously documented encoding_format=None regression for OpenAI-like embeddings.

---

Why I believe this is dependency-related

The current cocoindex-code pyproject.toml pins:

cocoindex[litellm]==1.0.0a38

So this issue appears to come from the cocoindex[litellm] / LiteLLM dependency path rather than from provider configuration alone.

Relevant source:

• cocoindex-code dependency declaration:
  https://github.com/cocoindex-io/cocoindex-code/blob/main/pyproject.toml

LiteLLM also published an official incident report on February 18, 2026 describing a very similar bug:

• a change explicitly sent encoding_format=None
• OpenAI-like embedding endpoints that strictly validated the field rejected the request
• the upstream fix was to filter out None / empty-string values before sending the request

Relevant source:

• LiteLLM incident report:
  https://docs.litellm.ai/blog/vllm-embeddings-incident

---

Environment

• OS: Windows
• Python: 3.12
• Installed via: uv tool install cocoindex-code

Embedding config:

embedding:
  provider: litellm
  model: openai/Qwen/Qwen3-Embedding-8B

envs:
  OPENAI_API_KEY: <redacted>
  OPENAI_BASE_URL: https://api.siliconflow.cn/v1

---

Reproduction

1. Configure cocoindex-code to use provider: litellm
2. Point it to an OpenAI-compatible embeddings endpoint
3. Run:

ccc index

4. Observe that indexing fails when the embeddings request is sent

---

Actual Behavior

The outgoing embeddings request contains:

{
  "input": ["hello"],
  "model": "Qwen/Qwen3-Embedding-8B",
  "encoding_format": null
}

The provider rejects this as an invalid parameter.

---

Expected Behavior

For OpenAI-compatible embedding providers:

• if encoding_format is unset, it should be omitted from the request body
• or it should be set to a valid value such as "float"

It should not be serialized as null.

---

What I found

I captured the outgoing request and confirmed that encoding_format: null was being sent.

I then patched my local LiteLLM installation as follows:

if encoding_format is not None:
    optional_params["encoding_format"] = encoding_format
else:
    optional_params["encoding_format"] = "float"

After this local patch, indexing worked.

A cleaner fix may be to omit None values entirely before sending the request to OpenAI-compatible embedding providers.

---

Why this seems to match the known LiteLLM bug

LiteLLM's official incident report says the regression was caused by explicitly passing encoding_format=None in embedding requests, and that the upstream fix was to filter out None and empty string values before sending requests to OpenAI-like providers.

That seems consistent with what I observed here:

• request fails when encoding_format: null is present
• request succeeds when the field is omitted
• request also succeeds when it is forced to "float"

---

Suggested Fix

One of the following would likely resolve the issue:

1. Upgrade the cocoindex[litellm] dependency path to a version that includes the upstream LiteLLM fix
2. Filter out encoding_format=None before sending embedding requests
3. As a fallback, explicitly send "float" instead of null

---

Notes

This does not appear to be model-specific.

The issue is the serialized request field:

"encoding_format": null

Once that field is omitted or changed to "float", the same provider/model combination works correctly.

[badmonster0]

Thanks a lot @lemoay , @georgeh0 can help with triage!

[wolpers]

Ahhh, wish I found this earlier. Here is a quick hack that would help with the described issue:
Fix in src/cocoindex_code/litellm_embedder.py:
line 59 — add encoding_format="float"
return await litellm.aembedding(model=model, input=input, encoding_format="float", **kwargs)