Embeddings — PentAGI

Embeddings power semantic search over memory, the knowledge base and past results. OpenAI is the default provider; with no EMBEDDING_MODEL set it uses text-embedding-ada-002. An example configuration in .env:

EMBEDDING_PROVIDER=openai          # openai, ollama, mistral, jina, huggingface, googleai, voyageai
EMBEDDING_MODEL=text-embedding-3-small
# optional custom endpoint
EMBEDDING_URL=
# provider API key, if required
EMBEDDING_KEY=
EMBEDDING_BATCH_SIZE=512               # default
EMBEDDING_MAX_TEXT_BYTES=8192          # default

Any service that implements OpenAI's /embeddings API works with EMBEDDING_PROVIDER=openai and a custom EMBEDDING_URL.

If you change the embedding provider or model, recompute the stored vectors with docker exec -it pentagi /opt/pentagi/bin/etester reindex — vectors from different models are not comparable. Do not run etester flush for this: it deletes every stored document, the knowledge base included.

Custom CA certificates

If a model backend or tool server uses a private CA and you see x509: certificate signed by unknown authority, put the PEM bundle into the pentagi-ssl volume and point PentAGI at it:

EXTERNAL_SSL_CA_PATH=/opt/pentagi/ssl/ca-bundle.pem
EXTERNAL_SSL_INSECURE=false

Then restart with docker compose up -d pentagi.