Embedding calls fail with OpenRouter: OpenAI SDK defaults to encoding_format=base64 which returns empty data

[merodahero]

Bug

When using EMBEDDING_MODEL_CONFIG__TRANSPORT=openai with an OpenAI-compatible provider that doesn't support encoding_format=base64 (e.g., OpenRouter hosting nvidia/nemotron-3-embed-1b:free), every embedding call fails with:

ValueError: No embedding data received

Root cause

The openai Python SDK (v2.36.0) defaults to encoding_format='base64' when the parameter is not explicitly passed:

# openai/resources/embeddings.py
if not is_given(encoding_format):
    params["encoding_format"] = "base64"

Honcho's src/embedding_client.py does not set encoding_format on either the single-query path (line 237) or the batch path (line 440), so the SDK sends base64 by default.

OpenRouter returns HTTP 200 but with empty embedding data for models that don't support base64 encoding. The SDK then raises ValueError("No embedding data received").

Reproduction

1. Configure Honcho with EMBEDDING_MODEL_CONFIG__MODEL=nvidia/nemotron-3-embed-1b:free on OpenRouter
2. Trigger any embedding (e.g., create_conclusions)
3. API logs show:

   WARNING - Embedding batch failed (attempt 1/3), retrying in 1s: No embedding data received
   ERROR - Error processing batch after all retries
   

Impact

• create_conclusions fails silently (returns error to caller)
• search_messages / semantic search fails
• Deriver fails to process queued observations
• Any model on OpenRouter (or other compatible providers) that doesn't support base64-encoded embeddings is affected

Fix

Add encoding_format='float' explicitly to both embedding call paths:

 # src/embedding_client.py line 237 (single-query)
-            openai_kwargs: dict[str, Any] = {"model": self.model, "input": [query]}
+            openai_kwargs: dict[str, Any] = {"model": self.model, "input": [query], "encoding_format": "float"}

 # src/embedding_client.py line 440 (batch)
                 openai_kwargs: dict[str, Any] = {
                     "model": self.model,
                     "input": [item.text for item in batch],
+                    "encoding_format": "float",
                 }

Environment

• Honcho v3.0.11 (commit 14538cf)
• openai Python SDK v2.36.0
• Provider: OpenRouter (https://openrouter.ai/api/v1)
• Model: nvidia/nemotron-3-embed-1b:free
• pgvector 0.8.2, column type halfvec(2048)

Workaround

Patch the file inside the container:

docker compose exec api sed -i \
  's/{"model": self.model, "input": \[query\]}/{"model": self.model, "input": [query], "encoding_format": "float"}/' \
  /app/src/embedding_client.py