Maintained by deepset

Integration: Ollama

Use Ollama models with Haystack. Ollama allows you to get up and running with large language models, locally.

Authors

Alistair Rogers

Sachin Sachdeva

deepset

GitHub Repo PyPI Package

Introduction
Installation
Usage
- Examples
  - Text Generation
  - Chat Generation

Introduction

You can use Ollama Models in your Haystack 2.0 pipelines with the OllamaGenerator.

Ollama is a project focused on running Large Language Models locally. Internally it uses the quantized GGUF format by default. This means it is possible to run LLMs on standard machines (even without GPUs), without having to handle complex installation procedures.

Installation

pip install ollama-haystack

Usage

This integration provides 2 components that allow you to leverage Ollama models:

The OllamaGenerator
The OllamaChatGenerator

To use an Ollama model:

Follow instructions on the Ollama Github Page to pull and serve your model of choice
Initialize one of the Ollama generators with the name of the model served in your Ollama instance.

Examples

To run the example, you may choose to run a docker container serving an Ollama model of your choice. Here are some commands that work with this example:

docker run -d -p 11434:11434 --name ollama ollama/ollama:latest
docker exec ollama ollama pull orca-mini

Text Generation

Below is the example of generative questions-answering pipeline using RAG with PromptBuilder and OllamaGenerator:

from haystack import Document, Pipeline
from haystack.components.builders.prompt_builder import PromptBuilder
from haystack.components.retrievers.in_memory import InMemoryBM25Retriever
from haystack.document_stores.in_memory import InMemoryDocumentStore

from haystack_integrations.components.generators.ollama import OllamaGenerator

document_store = InMemoryDocumentStore()
document_store.write_documents(
    [
        Document(content="Super Mario was an important politician"),
        Document(content="Mario owns several castles and uses them to conduct important political business"),
        Document(
            content="Super Mario was a successful military leader who fought off several invasion attempts by "
            "his arch rival - Bowser"
        ),
    ]
)

template = """
Given only the following information, answer the question.
Ignore your own knowledge.

Context:
{% for document in documents %}
    {{ document.content }}
{% endfor %}

Question: {{ query }}?
"""

pipe = Pipeline()

pipe.add_component("retriever", InMemoryBM25Retriever(document_store=document_store))
pipe.add_component("prompt_builder", PromptBuilder(template=template))
pipe.add_component("llm", OllamaGenerator(model="orca-mini", url="http://localhost:11434/api/generate"))
pipe.connect("retriever", "prompt_builder.documents")
pipe.connect("prompt_builder", "llm")

query = "Who is Super Mario?"

response = pipe.run({"prompt_builder": {"query": query}, "retriever": {"query": query}})

print(response["llm"]["replies"])

You should receive an output like (output is not deterministic):

['Based on the information provided, Super Mario is a successful military leader who fought
off several invasion attempts by his arch rival - Bowser. He is also an important politician and owns several
castles where he conducts political business. ' 'Therefore, it can be inferred that Super Mario is a combination of
both a military leader and an important politician.']

Chat Generation

from haystack.dataclasses import ChatMessage

from haystack_integrations.components.generators.ollama import OllamaChatGenerator

messages = [
    ChatMessage.from_user("What's Natural Language Processing?"),
    ChatMessage.from_system(
        "Natural Language Processing (NLP) is a field of computer science and artificial "
        "intelligence concerned with the interaction between computers and human language"
    ),
    ChatMessage.from_user("How do I get started?"),
]
client = OllamaChatGenerator(model="orca-mini", timeout=45, url="http://localhost:11434/api/chat")

response = client.run(messages, generation_kwargs={"temperature": 0.2})

print(response["replies"][0].content)