Local AI & Privacy

Ollama Integration

Configuring local open-weight AI via Ollama, prompt grounding boundaries, and ensuring hypotheses never override ground truth.

Last updated: 2026-10-31•v0.1.0

Last Signal integrates with locally hosted Ollama models to provide contextual root-cause reasoning without sacrificing data privacy. Unlike cloud AI tools, all prompts and model responses remain strictly on localhost.

Ollama Integration Architecture

Last Signal connects directly to the Ollama HTTP API at http://localhost:11434/api/generate. No external SDKs or telemetry wrappers are required.

Configuring Alternative Models

By default, Last Signal requests llama3.2 because of its optimal balance of fast local inference and structured reasoning. However, any model installed in your local Ollama instance can be specified:

bash

$# Use Qwen 2.5

$lastsignal report --model qwen2.5

$# Use DeepSeek Coder

$lastsignal report --model deepseek-coder:6.7b

$# Use Mistral

$lastsignal report --model mistral

Prompt Grounding Boundaries

To prevent LLM hallucination, Last Signal's prompt template enforces strict epistemic grounding. The model is provided only with:

The exact rule violations discovered by the deterministic engine.
The chronological sequence of failed operations and error strings.
The exit code and termination timestamp.

Hypothesis Invariant

The Ollama prompt explicitly instructs the model: 'You are an incident post-mortem assistant. Do not invent events that did not occur. State your conclusions strictly as hypotheses grounded in the provided evidence.'

Epistemic Discipline: Evidence vs. Hypotheses

In all generated post-mortem reports, outputs from Ollama are styled distinctly under the Local AI Hypothesis section with Mourning Violet visual indicators. They are never conflated with verified deterministic findings.