> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mezmo.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Ollama Guide

> Ollama configuration, fallback tool-call parsing, and practical guidance for local models.

AURA supports running local models through [Ollama](https://ollama.ai), including fallback parsing for tool-call formats that are emitted as plain text instead of native tool-call structures.

## Basic Configuration

```toml theme={null}
[agent]
name = "Local Assistant"
system_prompt = "You are a helpful assistant."

[agent.llm]
provider = "ollama"
model = "qwen3:30b-a3b"
# base_url = "http://localhost:11434" # optional; this default is used automatically
fallback_tool_parsing = true

[agent.llm.additional_params]
num_ctx = 32000
think   = true
```

`base_url` defaults to `http://localhost:11434` when omitted. Use `http://host.docker.internal:11434` when AURA runs inside a container and Ollama runs on the host.

All Ollama-specific parameters (`num_ctx`, `num_predict`, `think`, `seed`, `top_k`, `top_p`, etc.) go under `[agent.llm.additional_params]`. See [Ollama model parameters](https://github.com/ollama/ollama/blob/main/docs/modelfile.md#valid-parameters-and-values) for the full list.

## Fallback Tool Parsing

When `fallback_tool_parsing = true`, AURA tries to detect and execute tool calls from text output patterns commonly produced by local model families.

Known handled styles include:

* Pythonic-like calls (for example Llama-style patterns)
* XML-ish function wrappers (common in some Qwen outputs)
* JSON objects containing name/parameters payloads

This improves tool reliability with local models that do not consistently emit structured function-calling payloads.

<Warning>**Known issue (orchestration).** `fallback_tool_parsing` is currently honored only in single-agent mode. In orchestration mode the coordinator and workers are built with fallback parsing **disabled** — a bug, not intended behavior, tracked in [#193](https://github.com/mezmo/aura/issues/193). Until it's fixed, a local model that emits tool calls as text instead of native `tool_calls` can stall in orchestration (the coordinator never registers a routing decision). Workaround: run such models in single-agent mode (`[orchestration].enabled = false`), or use a model with reliable native tool-calling.</Warning>

## Practical Guidance

* Prefer instruction-tuned variants (`*-instruct`) when you need reliable tool execution.
* Keep prompts explicit about expected tool-call output format.
* Validate behavior with your exact model build and quantization.

## "Thinking model" Caveat

Thinking model variants have known malformed XML tool-call issues in some builds. AURA's fallback parser handles many of these cases, but reliability still depends on model artifact quality and prompt format constraints.