Consistency means producing predictable behavior across varied inputs and repeated interactions. In production, prompts act as quality contracts: they specify expected behavior and reduce ambiguity. Predictability supports reliability, but does not by itself ensure answers are correct or current. section 10.5.2 section 3.3.3
Choose the mechanism according to what must stay consistent:
- Use explicit prompts to specify expected behavior and escalation paths when a request exceeds the model’s capabilities. section 10.5.2
- Use RAG when answers should draw on one shared, updateable knowledge base; different phrasings can then draw on the same source. section 3.3.3
- Consider fine-tuning when detailed prompts fail to maintain a required style or format, or when specialized domain knowledge is needed. section 5.1.1
- For complex requests, specialized agents coordinated by an orchestrator can prevent one model from juggling unrelated responsibilities and producing inconsistent responses. section 8.0
A changing return policy
A support system can use RAG to answer differently worded questions from the same policy source. When the documents change, its answers can change too; consistency means sharing a source of truth, not freezing outdated answers. section 3.3.3
A required citation format
For a legal assistant that must use Bluebook citations, fine-tuning on correct examples is presented as a way to maintain the format in long conversations and respond cautiously when a case cannot be found. section 5.1.1
PROMPT_VERSIONS = { #1
"customer_support_v1.0": "You are a helpful customer support agent...",
"customer_support_v1.1": "You are a helpful customer support agent.
Always be empathetic...",
"customer_support_v2.0": "You are a customer support agent for AcmeSoft..."
}
def get_prompt(version="latest"): #2
if version == "latest":
version = max(PROMPT_VERSIONS.keys())
return PROMPT_VERSIONS[version]
Compare versioned customer-support prompts and see how a prompt can be retrieved by version. section 10.5.2
