[bug] Embedder silently defaults to OpenAI when EMBEDDING_* unset; no way to disable create_observations embedding

[PraveenKumarSridhar]

Summary

Two related defects in the embedding configuration path that affect anyone running self-hosted Honcho with a non-OpenAI LLM provider (Anthropic, MiniMax, Gemini, OpenRouter, local vLLM/Ollama, etc.).

(a) Silent fallback to OpenAI when EMBEDDING_* is unset. If no embedder config is provided, Honcho defaults EMBEDDING_MODEL_CONFIG to transport="openai", model="text-embedding-3-small", and routes to https://api.openai.com/v1/embeddings regardless of which LLM transport the rest of the stack uses. There is no startup-time warning that the embedder is pointing at OpenAI. A user running Honcho with LLM_OPENAI_BASE_URL=https://api.minimax.io/v1 (or any non-OpenAI LLM) will hit OpenAI 401s on every POST /v3/conclusions/ call unless they also configure an embedder. We hit this on a self-hosted stack at v3.0.7 and only discovered it because every conclusion write started 500-ing.

(b) No way to disable embedding on create_observations. EMBED_MESSAGES=false only gates retrieval-time embedding. There is no parallel EMBED_OBSERVATIONS flag and no transport="none" option. For self-hosted users who can't afford / don't want / can't reach an embedding API, every conclusion write either fails outright or requires a source-code patch to swallow the embedding error. The EmbeddingTransport Literal in src/config.py is ["openai", "gemini"] only — no path to a no-op.

Reproduction (a)

# .env: route LLM to a non-OpenAI provider, leave EMBEDDING_* unset
LLM_OPENAI_BASE_URL=https://api.minimax.io/v1
DERIVER_MODEL_CONFIG__TRANSPORT=openai
DERIVER_MODEL_CONFIG__MODEL=MiniMax-M3
DERIVER_MODEL_CONFIG__OVERRIDES__BASE_URL=https://api.minimax.io/v1

docker compose up -d
curl -X POST http://localhost:8000/v3/workspaces/test/peers/user -d '{"id":"user"}'
curl -X POST http://localhost:8000/v3/workspaces/test/conclusions \
  -d '{"conclusions":[{"observer_id":"ai","observed_id":"user","content":"hello"}]}' \
  -H "Content-Type: application/json"
# -> HTTP 500 from create_observations
# -> server log: openai.AuthenticationError: 401 ... sk-cp-... api.openai.com

Expected: at minimum, a startup warning that the embedder is using OpenAI defaults while the LLM is routed elsewhere. Ideally, inherit the LLM provider's transport/base_url/key when embedder config is unset.

Reproduction (b)

# User wants to disable embedding entirely (no budget, no provider, no GPU)
echo "EMBED_MESSAGES=false" >> .env
echo "EMBED_OBSERVATIONS=false" >> .env   # doesn't exist, ignored
docker compose restart honcho
# POST /v3/conclusions/ still calls simple_batch_embed(), still tries to hit
# the configured (or default) embedder, still fails the same way.

Expected: a documented way to disable embedding on create_observations. Either an EMBED_OBSERVATIONS=false flag, or a transport="none" option in EmbeddingTransport.

Environment

• Honcho v3.0.7 (self-hosted via docker-compose)
• LLM: MiniMax-M3 via OpenAI-compatible transport (DERIVER_MODEL_CONFIG__*)
• Embedder: left unset, defaults to OpenAI
• DB: PostgreSQL 16 with pgvector
• Reproduced on macOS 15.1, also confirmed by community on Raspberry Pi 4 / ARM64 with Ollama Cloud

Workarounds used

For (a), explicitly configure the embedder:

EMBEDDING_MODEL_CONFIG__TRANSPORT=openai
EMBEDDING_MODEL_CONFIG__MODEL=<embedding model>
EMBEDDING_MODEL_CONFIG__OVERRIDES__BASE_URL=<provider base url>
EMBEDDING_MODEL_CONFIG__OVERRIDES__API_KEY_ENV=LLM_OPENAI_API_KEY

For (b), source patch in src/embedding_client.py to swallow OpenAIAuthenticationError in simple_batch_embed and return zero-vectors. Documents save with NULL embedding; vector search is degraded but writes complete.

Adjacent issues considered but not duplicate

• Self-hosted Honcho with OpenAI-compatible providers (OpenRouter / vLLM): default AsyncOpenAI client lacks base_url → 401 against api.openai.com #641 (closed May 2026): LLM-side default AsyncOpenAI client lacks base_url, 401s against api.openai.com. Different layer (deriver LLM call, not embedder). Workaround was DERIVER_MODEL_CONFIG__* env vars.
• [Bug] Ollama/OpenAI-compatible providers: deriver silently produces zero observations (reasoning-content fallback + json_schema ignored) #716 (closed Jun 2026): Deriver silently produces zero observations on OpenAI-compatible providers. Different code path (Pydantic parse of response_format: json_schema).
• create_conclusion fails with local embedder due to dimension count mismatch #727 (open): create_conclusion fails with local embedder dimension mismatch. Related but distinct: dimension mismatch vs. silent fallback.
• Deriver silently marks representation work units processed when the embedding save fails on a transient 429 (no retry, no fail-loud → silent memory loss) #728 (open): Deriver silently marks representation processed when embedding save fails on transient 429. Related thematically (silent failure on embedding errors) but specific to retry behavior, not config defaults.
• OpenAI transport embeddings.create() missing dimensions parameter #625, embedding_client.py: OpenAI path doesn't send dimensions, silently ignoring EMBEDDING_VECTOR_DIMENSIONS #601, Allow configurable embedding model name + custom base URL for self-hosted embeddings #578, Feature Request: Configurable embedding model #554, Support local HTTP embedding providers with configurable base URL, model, and API key #518, Embedding provider parity: support 'custom' (OpenAI-compatible) in LLM_EMBEDDING_PROVIDER #513, [Bug] How to use a local embedding model ? #585 (closed): historical config/embedder-request issues, all addressed prior to v3.0.7.

Suggested fix

Either:

1. Inherit LLM provider config for embeddings when EMBEDDING_* is unset. If DERIVER_MODEL_CONFIG__OVERRIDES__BASE_URL is set, use that base URL for embeddings too. Same transport, same key.
2. Fail-fast at startup. If EMBEDDING_* is unset, log a clear "embedder using OpenAI defaults, set EMBEDDING_MODEL_CONFIG__* to override" warning at INFO or WARN level.
3. Add EMBED_OBSERVATIONS env flag parallel to EMBED_MESSAGES, default true.
4. Add transport="none" to EmbeddingTransport Literal as a documented no-op option.

Any of these would prevent the silent-401 trap. (1) + (3) is the cleanest pair.