handle_embeddings

akasha.helper.handle_embeddings() creates an embedding object from a provider alias, an existing embedding object, or a custom embedding function.

Create an embedding object

import akasha.helper as ah

embedding_model = ah.handle_embeddings(
    "openai:text-embedding-3-small"
)

The returned object follows the LangChain Embeddings interface.

Available methods

query_vector = embedding_model.embed_query("What is retrieval?")
document_vectors = embedding_model.embed_documents(
    ["Document one", "Document two"]
)

Method	Return value
embed_query(text)	One vector as a list of floating-point values.
embed_documents(texts)	A list of vectors, one for each input text.

Provider aliases

Common aliases include:

openai:text-embedding-3-small
gemini:gemini-embedding-001
ollama:nomic-embed-text
hf:BAAI/bge-base-en-v1.5

Ollama embeddings use the Ollama server and do not require a cloud API key. Pull the embedding model before using it:

ollama pull nomic-embed-text

Then create the embedding object with the ollama: alias:

import akasha.helper as ah

embedding_model = ah.handle_embeddings("ollama:nomic-embed-text")
query_vector = embedding_model.embed_query("What is retrieval?")

By default, Akasha connects to http://localhost:11434. Set OLLAMA_API_BASE to use another Ollama server:

OLLAMA_API_BASE=http://192.168.1.10:11434

You can also put the endpoint directly in the alias:

embedding_model = ah.handle_embeddings(
    "ollama:http://192.168.1.10:11434@nomic-embed-text"
)

Local Hugging Face embeddings may require the full installation. Other remote providers require their corresponding environment variables.

Note

The vector dimension is determined by the selected embedding model. Store and compare vectors produced by compatible models; vectors from different embedding spaces should not be mixed directly.