Local models

Every slot speaks the OpenAI wire shape, so anything that serves it works.

Ollama

ollama pull nomic-embed-text
ollama pull qwen2.5vl                   # a vision model, for scans
fluree config set doc.embedding.url http://localhost:11434/v1
fluree config set doc.embedding.model nomic-embed-text
fluree config set doc.vlm.url http://localhost:11434/v1
fluree config set doc.vlm.model qwen2.5vl

No key. Ollama’s /v1 route serves both embeddings and chat completions with images.

vLLM or LM Studio

Same shape: the server’s /v1 base as url, the served model name as model. LM Studio accepts any api_key; vLLM uses whatever you started it with.

OpenAI, or any hosted OpenAI-compatible API

fluree config set doc.embedding.url https://api.openai.com/v1
fluree config set doc.embedding.model text-embedding-3-small
fluree config set doc.embedding.api_key '$OPENAI_API_KEY'
fluree config set doc.vlm.url https://api.openai.com/v1
fluree config set doc.vlm.model gpt-5-mini
fluree config set doc.vlm.api_key '$OPENAI_API_KEY'

A $NAME value reads that environment variable at run time, so the key never sits in the config file. fluree config list redacts key values either way.

One model for both

Set only llm; vlm falls back to it. One multimodal model then reads crops and, when extraction is enabled, extracts entities.

Choosing an embedding size

Embeddings are stored at whatever width the model returns. Models that accept a dimensions parameter can be pinned with dimensions = 768 in the slot — smaller vectors are cheaper to store and to scan. Changing model or width later is fine: the documents are re-embedded because the model changed, and search reads the new vectors.

A query is only comparable to chunks embedded by the same model, so a corpus part-way through a model switch mixes widths. Re-ingest the whole folder after a change rather than a subset.