Replace hand-rolled provider functions with LiteLLM

[RheagalFire]

Motivation

prompt_models.py contains four separate provider functions (prompt_chatgpt, prompt_claude, prompt_deepseek, prompt_grok) plus three identical if/elif dispatcher chains that route by model name substring. Adding a new provider means writing a new function, updating all three dispatchers, and adding a new SDK dependency.

LiteLLM is an open-source AI gateway that provides a unified completion() interface for 100+ LLM providers using the OpenAI format. Replacing the hand-rolled functions with a single litellm.completion() call:

• Cuts prompt_models.py from 130 lines to 52
• Removes three copy-pasted if/elif dispatcher chains
• Makes adding a new provider zero-code (just add the model name to MODELS in workflow.py)
• Replaces direct openai + anthropic SDK dependencies with one dependency (litellm)

Changes

File	Change
prompt_orchestration/prompt_models.py	Replaced 4 provider functions + 3 dispatchers with single prompt_model() using litellm.completion()
requirements.txt	Removed openai==2.32.0 and anthropic==0.107.1, added litellm>=1.80.0,<1.87.0 (litellm installs both as transitive deps)

How provider routing works

LiteLLM routes based on model name prefixes (e.g., anthropic/claude-sonnet-4-6, bedrock/anthropic.claude-3-sonnet, vertex_ai/gemini-2.0-flash). A small _to_litellm_model() helper maps the repo's existing model names (deepseek-*, grok-*) to their LiteLLM prefix format so current configs keep working. Any of LiteLLM's 100+ supported providers can be used by passing the model name in LiteLLM format directly.

Env var compatibility

LiteLLM reads the standard env vars automatically:

Provider	Env var (same as before)
OpenAI	OPENAI_API_KEY
DeepSeek	DEEPSEEK_API_KEY
Anthropic	ANTHROPIC_API_KEY
xAI / Grok	XAI_API_KEY

For backward compatibility, if CLAUDE_API_KEY is set but ANTHROPIC_API_KEY is not, the module bridges them automatically.

Adding a new provider

Just add the model name in LiteLLM format to the MODELS list:

MODELS = [..., "vertex_ai/gemini-2.0-flash", "bedrock/anthropic.claude-3-sonnet"]

No new functions, no new SDK dependencies, no dispatcher changes.

Tested

$ python3 -c "
import litellm
response = litellm.completion(
    model='claude-sonnet-4-6',
    messages=[{'role': 'user', 'content': 'Say hello in exactly 5 words.'}],
    temperature=0.0, drop_params=True,
)
print(response.choices[0].message.content)
print(f'Model: {response.model}')
"
Hello there, how are you?
Model: claude-sonnet-4-6

Module import chain verified:

$ python3 -c "
from Experiments.multi_model_ipo.prompt_orchestration.prompt_models import (
    prompt_model, prompt_deep_research, prompt_daily_report, prompt_starting_report
)
print('All functions imported successfully')
"
All functions imported successfully