Skip to main content
AURA supports running local models through Ollama, including fallback parsing for tool-call formats that are emitted as plain text instead of native tool-call structures.

​
Basic Configuration

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 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.
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. 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.

​
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.