nucleusiq-ollama

Stable provider for running NucleusIQ agents against Ollama (local or hosted).

Install

0.2.2 — Development Status :: 5 - Production/Stable. Requires nucleusiq>=0.7.12 (pulled in automatically by pip).

pip install nucleusiq-ollama

Requires a running Ollama server unless you point OLLAMA_HOST at a remote instance. Optional: OLLAMA_API_KEY for authenticated endpoints.

Usage

from nucleusiq_ollama import BaseOllama, OllamaLLMParams

llm = BaseOllama(model_name="llama3.2", llm_params=OllamaLLMParams(think=True))

Runnable scripts (smoke, Agent DIRECT, streaming) live under examples/ — see examples/README.md.

See the Ollama provider guide for capability matrix, environment variables, and roadmap.

For Ollama's /v1 OpenAI shim (or any other Chat Completions server) use nucleusiq-openai-compatible instead — this package talks to the native Ollama API.

Status

0.2.2 — Development Status :: 5 - Production/Stable. Patch: multi-turn tool histories now send function.arguments as a mapping (the official ollama SDK rejects a JSON string). First stable line plus declared PROVIDER_NAME / dependency completeness.

• Chat + streaming + tools + structured outputs (JSON schema / format) + think pass-through.
• Vision (image messages) in the wire layer — OpenAI-style multimodal content lists with image_url parts whose URL is a data:image/*;base64,... data URL are converted to Ollama's message.images shape; HTTP(S) URLs are skipped with a warning (pre-encode them as data URLs to send images).
• Observability — every Ollama LLM call now lands in AgentResult.llm_calls with provider="ollama" populated automatically by the core agent loop.
• Floor nucleusiq>=0.7.12. 98 unit tests, coverage 99.85% (gate ≥ 95%).
• Embeddings + web search remain planned follow-ups for 0.2.x / 0.3.x.