Skip to content

Meet DeepintShield at Bengaluru Tech Summit 2026, 17 to 19 November, BIEC Hall 1, booth H1A106. Book 20 minutes

Start For Free

Get the same answer from an LLM every time

Reuse approved answers, pin golden answers for known questions and reduce randomness for the rest with Response Consistency.

The honest answer to how to get the same answer from an LLM every time is to stop regenerating answers you have already approved. DeepintShield Response Consistency reuses approved answers for matching questions, serves pinned golden answers from a versioned registry, and in Strict mode fixes temperature, seed and top-p for everything else.

The challenge

Language models sample their output, so the same question can produce different answers. For a creative tool that is fine. For policy, pricing or eligibility questions it is a problem:

  • Different customers, different answers. Two people ask the same refund question and get two versions of the rule.
  • Small changes, big effects. A reworded question or a model update shifts the answer.
  • Hard to review. Compliance teams cannot approve an answer that is generated fresh each time.
  • Hard to reproduce. When a complaint arrives, nobody can show what the assistant would have said.
  • Wasted generation. Producing the same approved answer again and again spends tokens for no benefit.

The approach

Separate questions with one correct answer from open-ended ones. For the first group, approve the answer once and reuse it. For the rest, reduce randomness and keep the model grounded in sources. Score a sample of answers so drift is visible. Fixed sampling settings reduce variation but do not replace reuse, and reuse applies to a narrower set of requests than all traffic, so test which of your routes qualify. Start with the questions that cause complaints or rework, such as refunds, fees or eligibility, and build the registry from answers your policy owner has approved. Decide who owns the registry and how often entries are reviewed, so approved answers do not go stale when the policy changes.

How to get the same answer from an LLM every time with DeepintShield

  1. Pick a consistency mode. Choose Off, Exact-Match, Semantic, Semantic + Pinned or Strict. Strict sets temperature 0, seed 42 and top-p 1. The semantic match threshold defaults to 0.93 (range 0.80 to 0.99), reused answers live for 72 hours by default, and a verifier can run off, on borderline matches or always. See the Hallucination Defense and Response Consistency docs.
  2. Pin golden answers. Load approved answers into the Golden Registry from CSV. Entries are versioned and scanned by guardrails, and the registry fails closed. Use it for the questions your compliance or support team has already signed off.
  3. Ground everything else. Turn on the proactive techniques: Stay grounded, Don’t make things up, Require sources, Hedge when unsure and Reduce randomness, which caps temperature at 0.4 by default. Strictness can be low, medium or high, with no extra model calls. Score answers after the response with five detectors (faithfulness, answer relevance, coherence, helpfulness and citation precision) combined into an overall hallucination-risk score; a 10% sample rate is recommended.

Hallucination Defense and Response Consistency are available on the Business plan and above.

What you can show afterwards

  • Which answers came from the Golden Registry, which were reused and which were generated.
  • The registry version behind each approved answer.
  • Questions that fell back to generation, which are candidates for the registry.
  • Hallucination-risk scores on sampled answers, and how they change after a model update.
  • The settings in force for each route, ready for a change review.

To test it, ask each registry question in several ways and confirm the approved answer comes back. Ask a question outside the registry and check it is answered with the grounding settings. Then update a registry entry and confirm the new version is served.

Read how to keep chatbot answers tied to policy and how LLM hallucination checks work.

Start For Free