How Model-Specific Parameters Like Temperature and Top_P Are Configured in the Hiring-Agent Repository

The interviewstreet/hiring-agent repository centralizes LLM generation settings in a single MODEL_PARAMETERS dictionary inside prompt.py, mapping each model name to its specific temperature and top_p values that propagate through the evaluation pipeline before being applied to provider-specific API calls.

The hiring-agent codebase supports multiple LLM backends including Ollama and Gemini, requiring a flexible system to manage inference parameters across different model architectures. Understanding how to configure model-specific parameters like temperature and top_p allows you to control response randomness, creativity, and determinism when scoring resumes or conducting automated technical interviews.

Centralized Configuration in prompt.py

All generation hyperparameters are defined in prompt.py within a top-level dictionary called MODEL_PARAMETERS. This mapping associates each supported model string with a dictionary containing its specific temperature and top_p values.

According to the source code at lines 27-44, the configuration looks like this:

# prompt.py

MODEL_PARAMETERS = {
    "qwen3:1.7b": {"temperature": 0.0, "top_p": 0.9},
    "gemma3:1b": {"temperature": 0.0, "top_p": 0.9},
    "qwen3:4b": {"temperature": 0.1, "top_p": 0.4},
    "gemma3:4b": {"temperature": 0.1, "top_p": 0.9},
    # … (other Ollama and Gemini models)

}

Each entry uses a plain Python dict, making the configuration human-readable and easy to version control without touching business logic.

Default Model Selection

The active model is determined by the DEFAULT_MODEL constant, which reads from an environment variable or falls back to "gemma3:4b" as specified in lines 15-22 of prompt.py:

# prompt.py (excerpt)

import os

DEFAULT_MODEL = os.getenv("DEFAULT_MODEL", "gemma3:4b")

This separation of model selection from parameter definition allows operators to switch models via environment variables while retaining the correct temperature and sampling parameters.

Parameter Propagation Pipeline

When the system evaluates a resume, the parameters flow through three distinct layers:

1. Entry Point (score.py) – Retrieves the parameter dictionary for the current default model and passes it to the evaluator:

   # score.py (excerpt)
   
   from prompt import DEFAULT_MODEL, MODEL_PARAMETERS
   
   model_params = MODEL_PARAMETERS.get(DEFAULT_MODEL)
   evaluator = ResumeEvaluator(
       model_name=DEFAULT_MODEL,
       model_params=model_params
   )

2. Orchestration (evaluator.py) – The ResumeEvaluator class stores these parameters and forwards them during the actual LLM invocation.

3. Provider Application (models.py) – At lines 40-46, the low-level provider implementation extracts temperature and top_p from the options dict and injects them into the provider-specific payload:

   # models.py (excerpt)
   
   generation_config = {}
   if "temperature" in options:
       generation_config["temperature"] = options["temperature"]
   if "top_p" in options:
       generation_config["top_p"] = options["top_p"]

   For Ollama models, these values are passed directly to client.chat(). For Gemini models, they become part of the generation_config dictionary when constructing the GenerativeModel instance.

Runtime Flexibility and Customization

Because MODEL_PARAMETERS is a plain dictionary, adding support for a new model requires only editing prompt.py. No changes are needed in score.py, evaluator.py, or models.py to recognize new parameter sets.

To retrieve parameters programmatically for inspection or logging:

from prompt import DEFAULT_MODEL, MODEL_PARAMETERS

params = MODEL_PARAMETERS.get(DEFAULT_MODEL, {})
print(f"Using model {DEFAULT_MODEL} with temperature={params.get('temperature')} "
      f"and top_p={params.get('top_p')}")

This outputs the active configuration (for example, when using the default "gemma3:4b"):

Using model gemma3:4b with temperature=0.1 and top_p=0.9

Summary

• prompt.py acts as the single source of truth, housing the MODEL_PARAMETERS dictionary that maps model names to their specific temperature and top_p values.
• Default model selection relies on the DEFAULT_MODEL environment variable, falling back to "gemma3:4b" when unset.
• Propagation flow moves parameters from score.py through evaluator.py and finally into models.py, where they are translated into provider-specific API calls.
• Extensibility is achieved through simple dictionary edits; the pipeline automatically picks up new model configurations without code changes elsewhere.

Frequently Asked Questions

How do I change the temperature for a specific model?

Modify the MODEL_PARAMETERS dictionary in prompt.py. Locate your model name (for example, "qwen3:4b") and update the "temperature" value. Because this is a centralized configuration, the change applies immediately to all subsequent evaluations using that model.

What happens if a model is not listed in MODEL_PARAMETERS?

If MODEL_PARAMETERS.get(DEFAULT_MODEL) returns None, downstream components receive an empty dictionary. In models.py, the absence of these keys means the LLM provider will use its own defaults, which may result in non-deterministic or overly creative outputs depending on the backend.

Can I use different parameters for Ollama versus Gemini models?

Yes. The MODEL_PARAMETERS dictionary can include entries for any model identifier string, regardless of provider. The provider-specific logic in models.py handles the translation of the generic temperature and top_p keys into the appropriate API format for each backend, allowing you to tune parameters per model architecture.

Is it possible to override parameters at runtime without editing prompt.py?

While the standard pipeline uses the static MODEL_PARAMETERS mapping, you can instantiate ResumeEvaluator directly with a custom model_params dictionary. Pass your override values when constructing the evaluator in score.py or your own entry script to temporarily adjust behavior without modifying the global configuration.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}