# Reward Model Service
== Schema ==
Message:
  # Same as /chat_completion

Options:
  logprobs: bool
  max_tokens: int
  temperature: float
  top_p: float
  #TODO: Figure out what other reward model specific params

ScoringFunction:
  name: str
  params: json

BatchRewardModelScoringJob:
  job_id: str
  created: str  # format date-time
  status: string  # enum (validating, running, completed, failed)
  input_file_path: Path
  success_file_path: Path  # jsonl where each row has {custom_id: <from input>, logprob: <float>, score: float}
  error_file_path: Path  # jsonl where each row has {custom_id: <from input>, error: <error_info>}
  metadata: json  # carry forward from job submission api


== Callsites ==

callsite:
  reward_model_scoring/
request_type:
  post
description:
  Score a prompt-response pair using a reward model
request:
  prompt_with_response: List[Message]  # prompt and response joined as a List[Message]
  reward_model: str
  scoring_function: ScoringFunction
  options: Options
response:
  id: str
  logprob: float
  score: float


callsite:
  reward_model_scoring/submit_job/
request_type:
  post
description:
  Batch scoring using reward models
request:
  reward_model: str
  prompt_with_response_path: Path  # jsonl file where each line is a List[Message] and custom_id
  scoring_function: ScoringFunction
  metadata: json  # anything to carry forward over in the response
response:
  batch_reward_model_scoring_job: BatchRewardModelScoringJob


callsite:
  /reward_model_scoring/submit_job/job_status
request_type:
  get
description:
  Get status for an already submitted job
request:
  job_id: str  # unique identifier for the job
response:
  batch_reward_model_scoring_job: BatchRewardModelScoringJob