DeepSeek-R1 / Reasoning Model Breaks JSON Parse or Enters Infinite Loop

Reasoning models emit chain-of-thought inside <think>...</think> before the final payload. Furthermore, forcing `"format": "json"` at token #0 suppresses the <think> phase and severely degrades reasoning accuracy.

1. Allow the model to emit its natural <think>...</think> block without grammar-level `"format": "json"`.
2. Strip everything up to `</think>` and optional markdown fences before calling JSON.parse().

export function parseReasoningJson(raw: string) {
  const withoutThink = raw.replace(/<think>[\s\S]*?<\/think>/gi, '').trim();
  const cleanJson = withoutThink
    .replace(/^\s*```(?:json)?\s*/i, '')
    .replace(/\s*```\s*$/i, '')
    .trim();
  return JSON.parse(cleanJson);
}

[BOUNDARY]If you only need fast deterministic extraction without chain-of-thought, use Qwen2.5-Instruct instead of DeepSeek-R1.

As documented in the official DeepSeek-R1 technical report, setting `temperature: 0` (greedy decoding) on quantized R1 distill models frequently traps the attention heads in self-doubt repetition loops.

1. Set `temperature: 0.6` (within the 0.5–0.7 sweet spot) and `top_p: 0.95`.
2. Add `repeat_penalty: 1.05` if using Q4_K_M quantization on long prompts.

{
  "model": "deepseek-r1:14b",
  "options": {
    "temperature": 0.6,
    "top_p": 0.95,
    "repeat_penalty": 1.05,
    "num_ctx": 16384
  }
}

[BOUNDARY]Also avoid adding a separate system prompt to R1; place all instructions inside the user message.

`max_tokens` (or `num_predict`) counts BOTH the `<think>` tokens and the final answer tokens. If `max_tokens` is left at 512 or 1024, the model spends all 1024 tokens thinking and truncates before `</think>`.

1. Set `num_predict: 4096` (or `max_tokens: 4096`) and `num_ctx: 16384` when calling DeepSeek-R1.
2. Check that the response contains a closing `</think>` tag before passing it downstream.

if (rawResponse.includes('<think>') && !rawResponse.includes('</think>')) {
  throw new Error('Reasoning model exhausted max_tokens inside <think> block before producing final answer');
}

[BOUNDARY]A typical medium-complexity coding or triage prompt consumes 600–1,800 <think> tokens before emitting the final answer.