How to generate Ollama embeddings in LangChain

Local embeddings let a retrieval application turn text into vectors without sending source material to a hosted API. In LangChain, the Ollama embedding integration calls a local Ollama server and returns vector lists that can later feed retrievers and vector stores.

The Python wrapper lives in the separate langchain-ollama package. Ollama still needs an embedding-capable model loaded locally; chat-only models can answer prompts but may fail when the embed endpoint asks for numeric vectors.

The smoke test uses all-minilm because it is a small embedding model and prints the dimensions for one query plus a two-document batch. Matching dimensions confirm that LangChain reached Ollama and received vectors from the selected model before the same object is used in a retriever or vector store.

1. Open an activated Python project environment.

   Related: How to create a virtual environment for LangChain

2. Pull the small Ollama embedding model.

   $ ollama pull all-minilm
   success

   all-minilm is an embedding-only model. Use another pulled embedding model when retrieval quality, language coverage, or vector size requirements differ.

3. Install the LangChain Ollama integration.

   $ python3 -m pip install --upgrade langchain-ollama

   langchain-ollama provides OllamaEmbeddings and uses the local Ollama API by default.
   Related: How to install LangChain with pip

4. Create the embedding smoke-test script.

   $ cat > langchain-ollama-embedding.py <<'PY'
   import os
   
   from langchain_ollama import OllamaEmbeddings
   
   
   model = os.environ.get("OLLAMA_EMBEDDING_MODEL", "all-minilm")
   embeddings = OllamaEmbeddings(
       model=model,
       base_url=os.environ.get("OLLAMA_BASE_URL", "http://localhost:11434"),
   )
   
   texts = [
       "Password reset tickets route to the identity queue.",
       "Invoice export tickets route to the billing queue.",
   ]
   
   query_vector = embeddings.embed_query("Where should password reset tickets go?")
   document_vectors = embeddings.embed_documents(texts)
   
   print(f"model: {model}")
   print(f"query dimensions: {len(query_vector)}")
   print(f"documents encoded: {len(document_vectors)}")
   print(f"first document dimensions: {len(document_vectors[0])}")
   print(f"first values: {[round(value, 4) for value in query_vector[:3]]}")
   PY

   embed_query() sends one search string. embed_documents() sends a list of document strings and should return one vector per document.

5. Run the embedding smoke test.

   $ python3 langchain-ollama-embedding.py
   model: all-minilm
   query dimensions: 384
   documents encoded: 2
   first document dimensions: 384
   first values: [0.005, -0.0435, -0.0524]

   The dimension count comes from the selected model. Rebuild stored vectors with the same embedding model and dimensions that the retriever will use.

6. Remove the temporary script.

   $ rm langchain-ollama-embedding.py