Semantic workflow

Local FastEmbed

uv tool install --reinstall --python 3.12 "ragvidence[semantic]==0.2.1"
ragvi semantic ./docs --query "How do I reset the device?"

First use downloads the selected model; inference then runs locally. --cache-dir chooses model/vector storage. Reuse requires compatible model and searchable-text identities. Dense retrieval ranks by cosine; weighted hybrid combines lexical/dense ranks with RRF.

Ollama embeddings

RAGVIDENCE_EMBEDDING_BASE_URL=http://localhost:11434

ragvi semantic ./docs --env-file ./providers.env --provider ollama --model YOUR_EMBEDDING_MODEL --query "Reset procedure"

Make a suitable embedding model available on your own Ollama service first. The CLI does not install a model server. Portable commands read dotenv only through explicit --env-file; they do not discover .env in your data folder.

OpenAI-compatible embeddings

RAGVIDENCE_EMBEDDING_BASE_URL=https://provider.example/v1
RAGVIDENCE_EMBEDDING_API_KEY=YOUR_KEY

Select --provider openai_compatible --model YOUR_EMBEDDING_MODEL and --env-file ./providers.env. Replace the example endpoint with your service. Compatible services can be local or remote; remote endpoints receive the source/query text needed for the operation. Keys are excluded from reports.

Compare specific models and settings

Hybrid, reranking and cosine MMR have different requirements. Use ragvi techniques before selecting them. Model output compliance and useful retrieval depend on the selected model and hardware; the CLI provides no API credits.

• Provider setup and installation