Temperature controls how adventurous token selection is during LLM generation. At 0, generation favors the most probable next token, producing highly consistent responses; higher values increase variety while potentially reducing coherence. Because temperature changes output variability, it also changes how production systems must test and evaluate responses. section 12.1.1
Choose temperature according to the operational behavior your application needs:
- Use temperature=0 when consistency, debugging, and reliable semantic evaluation matter most. Outputs are near-deterministic, but implementation details can still prevent exact reproducibility. section 12.2.6 section 12.5
- Use higher values such as 0.7–1.0 to increase response variety, while accepting a greater risk of reduced coherence. section 12.1.1
- Replace exact-string assertions with semantic correctness, factual accuracy, and guideline adherence when sampling makes repeated outputs differ. section 12.1.1
Consistent Gemini generation
A production RAG application can configure Gemini through LangChain with temperature=0. This favors the most probable tokens and supports consistent testing and debugging, while remaining short of guaranteed reproducibility. section 12.2.6 section 12.5
Listing 12.1 Parameters that control output randomness
temperature=0.0 #1 temperature=0.7 #2 temperature=1.5 #3 top_k=1 #4 top_k=40 #5 top_k=100 #6 top_p=0.1 #7 top_p=0.9 #8 top_p=1.0 #9
Compare temperature values from 0.0 through 1.5 alongside top-k and top-p settings that also control output randomness. section 12.1.1
Listing 12.11 Gemini config with zero temperature
llm = ChatGoogleGenerativeAI(
model="gemini-2.5-flash", #1
temperature=0., #2
google_api_key=os.environ["GEMINI_API_KEY"], #3
convert_system_message_to_human=True #4
)
Inspect a complete Gemini 2.5 Flash configuration using zero temperature for production-oriented consistency. section 12.2.6
