n8n / Automation Workflow Local LLM Node Failures

When an RSS feed returns 30 items, n8n fires requests faster than a single local GPU can process. Queued requests exceed the node timeout before inference starts.

1. Insert a Loop Over Items (Split in Batches) node with Batch Size = 1 before the local LLM node.
2. Set HTTP Request Timeout to 45000 ms and enable Retry On Fail (Max Tries: 3, Wait: 3000 ms).

// Recommended n8n HTTP Request Node Options for Local LLMs
{
  "batching": { "batch": { "batchSize": 1, "batchInterval": 500 } },
  "timeout": 45000,
  "retryOnFail": true,
  "maxTries": 3,
  "waitBetweenTries": 3000
}

[BOUNDARY]Matching batch size to OLLAMA_NUM_PARALLEL prevents VRAM thrashing and queue timeouts.

Asking for JSON only in the prompt text is not deterministic; small quantized models occasionally add conversational preambles or markdown code fences.

1. Set `"format": "json"` (or JSON Schema in Ollama 0.5+) at the grammar sampler level.
2. Enable `continueOnFail: true` so a single malformed item routes to an error branch instead of halting the whole batch.

// Ollama grammar-constrained JSON request body
{
  "model": "qwen2.5:7b",
  "format": "json",
  "stream": false,
  "options": { "temperature": 0.1 },
  "prompt": "Return JSON with keys {category, summary}: ..."
}

[BOUNDARY]Important: When format: "json" is enabled, you MUST also mention the word "JSON" in the prompt text.