hermes - 💡(How to fix) Fix Feature Request: Configurable Temperature Parameter for Model Inference [1 comments, 1 participants]

Problem

Hermes Agent currently does not expose a user-facing configuration for the LLM inference temperature parameter. The temperature is hardcoded internally via _fixed_temperature_for_model() or falls back to the API provider's default (typically 1.0). This causes severe hallucinations on certain models such as Minimax, where the default temperature leads to factually incorrect outputs.

Current Workarounds (Insufficient)

• Internal code path: request_overrides can set temperature, but this is not exposed via config.yaml or CLI
• Kimi models: hardcoded to omit temperature entirely (special case)
• Other models: no temperature sent → relies on provider default

No user-configurable option exists in:

• config.yaml (model: section)
• hermes chat CLI flags
• Environment variables

Proposed Solution

Implement a configuration exposure via config.yaml and CLI. Underlying transport already supports fixed_temperature — this only exposes it to users.

Technical Feasibility

• ChatCompletionsTransport.build_kwargs() accepts fixed_temperature parameter
• _fixed_temperature_for_model() can be extended to read config
• No core inference changes required
• Backward compatible: default behavior unchanged

Request

Add temperature configuration in model.temperature (config.yaml) and --temperature (CLI).