Embeddings and costs

Embeddings are optional

Ceres does not require embeddings to be useful.

You can run the full harvesting pipeline in metadata-only mode, keep a synchronized catalog in PostgreSQL, export it, and operate the API without ever generating vectors.

Recommended path: local Ollama

If you want semantic search, the preferred setup is local Ollama:

ollama serve
ollama pull nomic-embed-text

export EMBEDDING_PROVIDER=ollama
export OLLAMA_ENDPOINT=http://localhost:11434
ceres embed

That gives you:

• zero per-request embedding cost
• a clean separation between harvesting and embedding runs
• the option to backfill vectors only when you decide the catalog is ready

When cloud providers still make sense

Gemini and OpenAI are still supported when you want managed infrastructure or different model characteristics. In that case, your cost is driven by the standalone embedding pass, not by harvesting itself.

Typical pattern:

1. Harvest with --metadata-only
2. Decide whether a portal or snapshot is worth embedding
3. Run ceres embed with your chosen provider

That keeps cloud spend explicit and isolated.

Cost shape by workflow

Metric	Detail
Harvest only	No embedding cost
Harvest + local Ollama	No per-request cost, only your own CPU/GPU/runtime cost
Harvest + Gemini/OpenAI	Pay only for the texts you choose to embed
Re-sync after first run	Usually cheap because incremental sync and delta detection reduce re-embedding

Why later runs are cheaper

Ceres uses two mechanisms to keep optional embedding passes under control:

• incremental sync reduces the amount of portal data fetched
• content-hash delta detection skips re-embedding when the embeddable text did not actually change

That matters whether you run Ollama locally or use a hosted provider.

Practical guidance

• If your goal is catalog maintenance, stay in metadata-only mode.
• If your goal is semantic retrieval, start with Ollama.
• If you need hosted inference, harvest first and embed second so the spend is isolated and reversible.

The important shift is operational: in Ceres, embeddings are now an optional downstream stage, not a prerequisite for using the project.