OpenAI Embeddings

Historical evaluation of the architecture before Topic removal. The measurements below are retained for comparison and do not describe current processing.

OpenAI is the default embedding provider. RSSMonster uses text-embedding-3-small with 1536 dimensions unless it is configured otherwise. Embedding requests still pass through the local inference service; only that service contacts OpenAI.

Configuration

Configure the RSSMonster server to reach inference:

# server/.env
INFERENCE_BASE_URL=http://127.0.0.1:3001
INFERENCE_TIMEOUT_MS=30000
# Optional assistant reasoning override; use only a value supported by the model.
ASSISTANT_REASONING_EFFORT=

Configure each remote capability in inference/.env:

EMBEDDING_PROVIDER=openai-compatible
EMBEDDING_BASE_URL=https://api.openai.com/v1
EMBEDDING_API_KEY=your-embedding-api-key
EMBEDDING_MODEL=text-embedding-3-small
EMBEDDING_DIMENSIONS=1536

GENERATION_PROVIDER=openai-compatible
GENERATION_BASE_URL=https://api.openai.com/v1
GENERATION_API_KEY=your-generation-api-key
GENERATION_MODEL=gpt-4o-mini

CLASSIFICATION_PROVIDER=openai-compatible
CLASSIFICATION_BASE_URL=https://api.openai.com/v1
CLASSIFICATION_API_KEY=your-classification-api-key
CLASSIFICATION_MODEL=gpt-4o-mini

ASSISTANT_PROVIDER=openai-compatible
ASSISTANT_BASE_URL=https://api.openai.com/v1
ASSISTANT_API_KEY=your-assistant-api-key
ASSISTANT_MODEL=gpt-4o-mini

OpenAI uses the generic openai-compatible adapter. Each capability can instead use a different compatible endpoint and credential. Remote classification uses the existing LLM scoring flow. The assistant uses Chat Completions through the Agents SDK and requires tool calling and streaming support. No local model is loaded for a capability selected as external.

GENERATION_MODEL defaults all generated text workloads to one model; optional GENERATION_ARTICLE_MODEL, GENERATION_SMART_FOLDER_MODEL, and GENERATION_FEED_REDISCOVERY_MODEL override it. Set OPENAI_OMIT_TEMPERATURE=true when the endpoint rejects explicit temperature. Assistant reasoning is optional and belongs in server/.env.

See Inference for mixed endpoint examples, container networking, configuration validation, and legacy migration. Keep credentials exclusively in inference. Existing openai provider names and OPENAI_* model/connection settings remain deprecated aliases for one release; migrate providers and model settings together.

Observed Semantic Behavior

In RSSMonster’s 686-article semantic regression fixture, OpenAI passed the same 12 pipeline tests as Qwen. It generally produced more conservative event and topic grouping and fewer interest-island fallback assignments. Duplicate detection produced the same result with both models.

The report also showed trade-offs: OpenAI kept some related articles out of otherwise coherent events, and one topic incorrectly joined two unrelated public-transport accidents. It grouped some Windows and game-related article sets more completely than Qwen. These are observations from one RSSMonster fixture, not general guarantees about the model.

OpenAI avoids the local model download and CPU/memory cost, but embedding content is sent to an external API and incurs API usage. Its vector space is not compatible with Qwen vectors.