mirror of
https://github.com/BerriAI/litellm.git
synced 2026-10-12 03:48:56 +00:00
refactor(decisions): move the OpenAI wire translation into llms/openai
(cherry picked from commit d209a85076)
This commit is contained in:
mateo-berri
parent
138f901e09
commit
fbcc73b579
5 changed files with 227 additions and 203 deletions
|
|
@ -1,21 +1,14 @@
|
|||
import json
|
||||
from collections.abc import Mapping, Sequence
|
||||
from typing import Final
|
||||
|
||||
from pydantic import TypeAdapter
|
||||
from typing_extensions import assert_never
|
||||
|
||||
from litellm.types.decisions import (
|
||||
ChoiceAnswer,
|
||||
ChoiceQuestion,
|
||||
DecisionAnswer,
|
||||
DecisionQuestion,
|
||||
DecisionsJSON,
|
||||
DecisionsRequestBody,
|
||||
DecisionsResponse,
|
||||
DecisionsUsage,
|
||||
NoulAnswer,
|
||||
NoulQuestion,
|
||||
OpenAIChoiceAnswer,
|
||||
OpenAIChoiceProbability,
|
||||
OpenAIChoiceQuestion,
|
||||
|
|
@ -33,12 +26,10 @@ from litellm.types.decisions import (
|
|||
OpenAIScoreProbability,
|
||||
OpenAIScoreQuestion,
|
||||
ScoreAnswer,
|
||||
ScoreQuestion,
|
||||
systemone_choice_key,
|
||||
)
|
||||
|
||||
_OPENAI_ONLY_FIELDS: Final = frozenset({"input", "questions", "safety_identifier"})
|
||||
_OPENAI_RESPONSE_ADAPTER: Final[TypeAdapter[OpenAIDecisionResponse]] = TypeAdapter(OpenAIDecisionResponse)
|
||||
|
||||
|
||||
def _message_text(message: OpenAIDecisionInputMessage) -> str:
|
||||
|
|
@ -134,92 +125,3 @@ def to_openai_response(
|
|||
total_tokens=usage.input_tokens + usage.output_tokens,
|
||||
),
|
||||
)
|
||||
|
||||
|
||||
def _text(value: DecisionsJSON) -> str:
|
||||
return value if isinstance(value, str) else json.dumps(value)
|
||||
|
||||
|
||||
def _instructions(instructions: DecisionsJSON | None) -> Mapping[str, str]:
|
||||
return {} if instructions is None else {"instructions": _text(instructions)}
|
||||
|
||||
|
||||
def _predicate_instructions(question: NoulQuestion) -> str:
|
||||
instructions: Final = () if question.instructions is None else (_text(question.instructions),)
|
||||
criteria: Final = tuple(
|
||||
f"Answer {answer} when: {_text(rule)}" for answer, rule in (question.criteria or {}).items() if rule is not None
|
||||
)
|
||||
return "\n\n".join((*instructions, *criteria))
|
||||
|
||||
|
||||
def _choice(value: str, description: DecisionsJSON | None) -> Mapping[str, str]:
|
||||
return {"value": value} if description is None else {"value": value, "description": _text(description)}
|
||||
|
||||
|
||||
def _openai_question(name: str, question: DecisionQuestion) -> Mapping[str, object]:
|
||||
match question:
|
||||
case NoulQuestion():
|
||||
return {"type": "predicate", "name": name, "instructions": _predicate_instructions(question)}
|
||||
case ChoiceQuestion():
|
||||
return {
|
||||
"type": "choice",
|
||||
"name": name,
|
||||
**_instructions(question.instructions),
|
||||
"choices": [_choice(value, description) for value, description in question.criteria.items()],
|
||||
}
|
||||
case ScoreQuestion():
|
||||
return {
|
||||
"type": "score",
|
||||
"name": name,
|
||||
**_instructions(question.instructions),
|
||||
"levels": [{"label": _text(criterion)} for criterion in question.criteria],
|
||||
}
|
||||
case _:
|
||||
assert_never(question)
|
||||
|
||||
|
||||
def to_openai_request(model: str, request: DecisionsRequestBody) -> Mapping[str, object]:
|
||||
return {
|
||||
"model": model,
|
||||
"input": _text(request.state),
|
||||
"questions": [_openai_question(name, question) for name, question in request.questions.items()],
|
||||
}
|
||||
|
||||
|
||||
def _systemone_answer(answer: OpenAIDecisionAnswer) -> DecisionAnswer | None:
|
||||
match answer:
|
||||
case OpenAIPredicateAnswer():
|
||||
return NoulAnswer(type="noul", noul=answer.probability)
|
||||
case OpenAIChoiceAnswer():
|
||||
return ChoiceAnswer(
|
||||
type="choice",
|
||||
choice=systemone_choice_key(answer.choice),
|
||||
confidence=answer.confidence,
|
||||
probabilities={systemone_choice_key(item.value): item.probability for item in answer.probabilities},
|
||||
)
|
||||
case OpenAIScoreAnswer():
|
||||
return ScoreAnswer(
|
||||
type="score",
|
||||
score=answer.score,
|
||||
confidence=answer.confidence,
|
||||
legend={str(item.value): item.label for item in answer.probabilities},
|
||||
probabilities={str(item.value): item.probability for item in answer.probabilities},
|
||||
)
|
||||
case OpenAIRefusalAnswer():
|
||||
return None
|
||||
case _:
|
||||
assert_never(answer)
|
||||
|
||||
|
||||
def _named_answers(answers: Sequence[OpenAIDecisionAnswer]) -> Mapping[str, DecisionAnswer]:
|
||||
translated: Final = ((answer.name, _systemone_answer(answer)) for answer in answers)
|
||||
return {name: answer for name, answer in translated if name is not None and answer is not None}
|
||||
|
||||
|
||||
def to_systemone_response(payload: object) -> DecisionsResponse:
|
||||
response: Final = _OPENAI_RESPONSE_ADAPTER.validate_python(payload)
|
||||
return DecisionsResponse(
|
||||
model=response.model,
|
||||
answers=_named_answers(response.answers),
|
||||
usage=DecisionsUsage(input_tokens=response.usage.input_tokens, output_tokens=response.usage.output_tokens),
|
||||
)
|
||||
|
|
|
|||
|
|
@ -1,9 +1,123 @@
|
|||
from collections.abc import Mapping
|
||||
import json
|
||||
from collections.abc import Mapping, Sequence
|
||||
from dataclasses import dataclass
|
||||
from typing import Final
|
||||
|
||||
from litellm.decisions.openai_transformation import to_openai_request, to_systemone_response
|
||||
from litellm.types.decisions import DecisionsRequestBody, DecisionsResponse
|
||||
from pydantic import TypeAdapter
|
||||
from typing_extensions import assert_never
|
||||
|
||||
from litellm.types.decisions import (
|
||||
ChoiceAnswer,
|
||||
ChoiceQuestion,
|
||||
DecisionAnswer,
|
||||
DecisionQuestion,
|
||||
DecisionsJSON,
|
||||
DecisionsRequestBody,
|
||||
DecisionsResponse,
|
||||
DecisionsUsage,
|
||||
NoulAnswer,
|
||||
NoulQuestion,
|
||||
OpenAIChoiceAnswer,
|
||||
OpenAIDecisionAnswer,
|
||||
OpenAIDecisionResponse,
|
||||
OpenAIPredicateAnswer,
|
||||
OpenAIRefusalAnswer,
|
||||
OpenAIScoreAnswer,
|
||||
ScoreAnswer,
|
||||
ScoreQuestion,
|
||||
systemone_choice_key,
|
||||
)
|
||||
|
||||
_OPENAI_RESPONSE_ADAPTER: Final[TypeAdapter[OpenAIDecisionResponse]] = TypeAdapter(OpenAIDecisionResponse)
|
||||
|
||||
|
||||
def _text(value: DecisionsJSON) -> str:
|
||||
return value if isinstance(value, str) else json.dumps(value)
|
||||
|
||||
|
||||
def _instructions(instructions: DecisionsJSON | None) -> Mapping[str, str]:
|
||||
return {} if instructions is None else {"instructions": _text(instructions)}
|
||||
|
||||
|
||||
def _predicate_instructions(question: NoulQuestion) -> str:
|
||||
instructions: Final = () if question.instructions is None else (_text(question.instructions),)
|
||||
criteria: Final = tuple(
|
||||
f"Answer {answer} when: {_text(rule)}" for answer, rule in (question.criteria or {}).items() if rule is not None
|
||||
)
|
||||
return "\n\n".join((*instructions, *criteria))
|
||||
|
||||
|
||||
def _choice(value: str, description: DecisionsJSON | None) -> Mapping[str, str]:
|
||||
return {"value": value} if description is None else {"value": value, "description": _text(description)}
|
||||
|
||||
|
||||
def _openai_question(name: str, question: DecisionQuestion) -> Mapping[str, object]:
|
||||
match question:
|
||||
case NoulQuestion():
|
||||
return {"type": "predicate", "name": name, "instructions": _predicate_instructions(question)}
|
||||
case ChoiceQuestion():
|
||||
return {
|
||||
"type": "choice",
|
||||
"name": name,
|
||||
**_instructions(question.instructions),
|
||||
"choices": [_choice(value, description) for value, description in question.criteria.items()],
|
||||
}
|
||||
case ScoreQuestion():
|
||||
return {
|
||||
"type": "score",
|
||||
"name": name,
|
||||
**_instructions(question.instructions),
|
||||
"levels": [{"label": _text(criterion)} for criterion in question.criteria],
|
||||
}
|
||||
case _:
|
||||
assert_never(question)
|
||||
|
||||
|
||||
def to_openai_request(model: str, request: DecisionsRequestBody) -> Mapping[str, object]:
|
||||
return {
|
||||
"model": model,
|
||||
"input": _text(request.state),
|
||||
"questions": [_openai_question(name, question) for name, question in request.questions.items()],
|
||||
}
|
||||
|
||||
|
||||
def _systemone_answer(answer: OpenAIDecisionAnswer) -> DecisionAnswer | None:
|
||||
match answer:
|
||||
case OpenAIPredicateAnswer():
|
||||
return NoulAnswer(type="noul", noul=answer.probability)
|
||||
case OpenAIChoiceAnswer():
|
||||
return ChoiceAnswer(
|
||||
type="choice",
|
||||
choice=systemone_choice_key(answer.choice),
|
||||
confidence=answer.confidence,
|
||||
probabilities={systemone_choice_key(item.value): item.probability for item in answer.probabilities},
|
||||
)
|
||||
case OpenAIScoreAnswer():
|
||||
return ScoreAnswer(
|
||||
type="score",
|
||||
score=answer.score,
|
||||
confidence=answer.confidence,
|
||||
legend={str(item.value): item.label for item in answer.probabilities},
|
||||
probabilities={str(item.value): item.probability for item in answer.probabilities},
|
||||
)
|
||||
case OpenAIRefusalAnswer():
|
||||
return None
|
||||
case _:
|
||||
assert_never(answer)
|
||||
|
||||
|
||||
def _named_answers(answers: Sequence[OpenAIDecisionAnswer]) -> Mapping[str, DecisionAnswer]:
|
||||
translated: Final = ((answer.name, _systemone_answer(answer)) for answer in answers)
|
||||
return {name: answer for name, answer in translated if name is not None and answer is not None}
|
||||
|
||||
|
||||
def to_systemone_response(payload: object) -> DecisionsResponse:
|
||||
response: Final = _OPENAI_RESPONSE_ADAPTER.validate_python(payload)
|
||||
return DecisionsResponse(
|
||||
model=response.model,
|
||||
answers=_named_answers(response.answers),
|
||||
usage=DecisionsUsage(input_tokens=response.usage.input_tokens, output_tokens=response.usage.output_tokens),
|
||||
)
|
||||
|
||||
|
||||
@dataclass(frozen=True, slots=True)
|
||||
|
|
|
|||
|
|
@ -4,7 +4,7 @@ from typing import Final
|
|||
import pytest
|
||||
from pydantic import TypeAdapter, ValidationError
|
||||
|
||||
from litellm.decisions.openai_transformation import to_openai_request, to_systemone_request, to_systemone_response
|
||||
from litellm.decisions.openai_transformation import to_systemone_request
|
||||
from litellm.types.decisions import MAX_DECISION_QUESTIONS, DecisionsRequestBody, OpenAIDecisionRequestBody
|
||||
|
||||
_OPENAI_BODY: Final[TypeAdapter[OpenAIDecisionRequestBody]] = TypeAdapter(OpenAIDecisionRequestBody)
|
||||
|
|
@ -30,104 +30,3 @@ def test_the_largest_openai_request_accepted_translates_to_a_valid_systemone_req
|
|||
def test_an_openai_request_with_more_questions_than_systemone_takes_is_rejected_before_translation() -> None:
|
||||
with pytest.raises(ValidationError, match="questions"):
|
||||
_OPENAI_BODY.validate_python(_openai_request(MAX_DECISION_QUESTIONS + 1))
|
||||
|
||||
|
||||
def test_a_systemone_request_becomes_an_openai_request_with_questions_named_by_their_keys() -> None:
|
||||
request: Final = _SYSTEMONE_BODY.validate_python(
|
||||
{
|
||||
"state": {"ticket": 1234, "text": "Screen cracked"},
|
||||
"questions": {
|
||||
"damaged": {
|
||||
"type": "noul",
|
||||
"instructions": "Is the item damaged?",
|
||||
"criteria": {"true": "Visible damage", "false": None},
|
||||
"provider_field": "dropped",
|
||||
},
|
||||
"rubric_only": {"type": "noul", "criteria": {"true": {"signal": "refund"}}},
|
||||
"action": {
|
||||
"type": "choice",
|
||||
"instructions": {"policy": "refund-v2"},
|
||||
"criteria": {"refund": "Within 30 days", "escalate": None},
|
||||
},
|
||||
"severity": {"type": "score", "criteria": ["minor", {"label": "major"}]},
|
||||
},
|
||||
}
|
||||
)
|
||||
|
||||
assert to_openai_request("gpt-6-luna", request) == {
|
||||
"model": "gpt-6-luna",
|
||||
"input": '{"ticket": 1234, "text": "Screen cracked"}',
|
||||
"questions": [
|
||||
{
|
||||
"type": "predicate",
|
||||
"name": "damaged",
|
||||
"instructions": "Is the item damaged?\n\nAnswer true when: Visible damage",
|
||||
},
|
||||
{"type": "predicate", "name": "rubric_only", "instructions": 'Answer true when: {"signal": "refund"}'},
|
||||
{
|
||||
"type": "choice",
|
||||
"name": "action",
|
||||
"instructions": '{"policy": "refund-v2"}',
|
||||
"choices": [{"value": "refund", "description": "Within 30 days"}, {"value": "escalate"}],
|
||||
},
|
||||
{"type": "score", "name": "severity", "levels": [{"label": "minor"}, {"label": '{"label": "major"}'}]},
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
def test_an_openai_response_becomes_systemone_answers_keyed_by_question_name_without_refusals() -> None:
|
||||
response: Final = to_systemone_response(
|
||||
{
|
||||
"model": "gpt-6-luna",
|
||||
"answers": [
|
||||
{"type": "predicate", "name": "damaged", "probability": 0.95},
|
||||
{
|
||||
"type": "choice",
|
||||
"name": "action",
|
||||
"choice": True,
|
||||
"probabilities": [{"value": True, "probability": 0.9}, {"value": "escalate", "probability": 0.1}],
|
||||
"confidence": 0.8,
|
||||
},
|
||||
{
|
||||
"type": "score",
|
||||
"name": "severity",
|
||||
"score": 1.72,
|
||||
"probabilities": [
|
||||
{"value": 0, "label": "minor", "probability": 0.0},
|
||||
{"value": 1, "label": "major", "probability": 0.28},
|
||||
{"value": 2, "label": "critical", "probability": 0.72},
|
||||
],
|
||||
"confidence": 0.58,
|
||||
},
|
||||
{"type": "refusal", "name": "fraud"},
|
||||
],
|
||||
"usage": {
|
||||
"input_tokens": 383,
|
||||
"input_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 0},
|
||||
"output_tokens": 2,
|
||||
"output_tokens_details": {"reasoning_tokens": 0},
|
||||
"total_tokens": 385,
|
||||
},
|
||||
}
|
||||
)
|
||||
|
||||
assert response.model_dump(mode="json") == {
|
||||
"model": "gpt-6-luna",
|
||||
"answers": {
|
||||
"damaged": {"type": "noul", "noul": 0.95},
|
||||
"action": {
|
||||
"type": "choice",
|
||||
"choice": "true",
|
||||
"confidence": 0.8,
|
||||
"probabilities": {"true": 0.9, "escalate": 0.1},
|
||||
},
|
||||
"severity": {
|
||||
"type": "score",
|
||||
"score": 1.72,
|
||||
"confidence": 0.58,
|
||||
"legend": {"0": "minor", "1": "major", "2": "critical"},
|
||||
"probabilities": {"0": 0.0, "1": 0.28, "2": 0.72},
|
||||
},
|
||||
},
|
||||
"usage": {"input_tokens": 383, "output_tokens": 2},
|
||||
}
|
||||
|
|
|
|||
0
tests/unit/llms/openai/decisions/__init__.py
Normal file
0
tests/unit/llms/openai/decisions/__init__.py
Normal file
|
|
@ -0,0 +1,109 @@
|
|||
from typing import Final
|
||||
|
||||
from pydantic import TypeAdapter
|
||||
|
||||
from litellm.llms.openai.decisions.transformation import to_openai_request, to_systemone_response
|
||||
from litellm.types.decisions import DecisionsRequestBody
|
||||
|
||||
_SYSTEMONE_BODY: Final[TypeAdapter[DecisionsRequestBody]] = TypeAdapter(DecisionsRequestBody)
|
||||
|
||||
|
||||
def test_a_systemone_request_becomes_an_openai_request_with_questions_named_by_their_keys() -> None:
|
||||
request: Final = _SYSTEMONE_BODY.validate_python(
|
||||
{
|
||||
"state": {"ticket": 1234, "text": "Screen cracked"},
|
||||
"questions": {
|
||||
"damaged": {
|
||||
"type": "noul",
|
||||
"instructions": "Is the item damaged?",
|
||||
"criteria": {"true": "Visible damage", "false": None},
|
||||
"provider_field": "dropped",
|
||||
},
|
||||
"rubric_only": {"type": "noul", "criteria": {"true": {"signal": "refund"}}},
|
||||
"action": {
|
||||
"type": "choice",
|
||||
"instructions": {"policy": "refund-v2"},
|
||||
"criteria": {"refund": "Within 30 days", "escalate": None},
|
||||
},
|
||||
"severity": {"type": "score", "criteria": ["minor", {"label": "major"}]},
|
||||
},
|
||||
}
|
||||
)
|
||||
|
||||
assert to_openai_request("gpt-6-luna", request) == {
|
||||
"model": "gpt-6-luna",
|
||||
"input": '{"ticket": 1234, "text": "Screen cracked"}',
|
||||
"questions": [
|
||||
{
|
||||
"type": "predicate",
|
||||
"name": "damaged",
|
||||
"instructions": "Is the item damaged?\n\nAnswer true when: Visible damage",
|
||||
},
|
||||
{"type": "predicate", "name": "rubric_only", "instructions": 'Answer true when: {"signal": "refund"}'},
|
||||
{
|
||||
"type": "choice",
|
||||
"name": "action",
|
||||
"instructions": '{"policy": "refund-v2"}',
|
||||
"choices": [{"value": "refund", "description": "Within 30 days"}, {"value": "escalate"}],
|
||||
},
|
||||
{"type": "score", "name": "severity", "levels": [{"label": "minor"}, {"label": '{"label": "major"}'}]},
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
def test_an_openai_response_becomes_systemone_answers_keyed_by_question_name_without_refusals() -> None:
|
||||
response: Final = to_systemone_response(
|
||||
{
|
||||
"model": "gpt-6-luna",
|
||||
"answers": [
|
||||
{"type": "predicate", "name": "damaged", "probability": 0.95},
|
||||
{
|
||||
"type": "choice",
|
||||
"name": "action",
|
||||
"choice": True,
|
||||
"probabilities": [{"value": True, "probability": 0.9}, {"value": "escalate", "probability": 0.1}],
|
||||
"confidence": 0.8,
|
||||
},
|
||||
{
|
||||
"type": "score",
|
||||
"name": "severity",
|
||||
"score": 1.72,
|
||||
"probabilities": [
|
||||
{"value": 0, "label": "minor", "probability": 0.0},
|
||||
{"value": 1, "label": "major", "probability": 0.28},
|
||||
{"value": 2, "label": "critical", "probability": 0.72},
|
||||
],
|
||||
"confidence": 0.58,
|
||||
},
|
||||
{"type": "refusal", "name": "fraud"},
|
||||
],
|
||||
"usage": {
|
||||
"input_tokens": 383,
|
||||
"input_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 0},
|
||||
"output_tokens": 2,
|
||||
"output_tokens_details": {"reasoning_tokens": 0},
|
||||
"total_tokens": 385,
|
||||
},
|
||||
}
|
||||
)
|
||||
|
||||
assert response.model_dump(mode="json") == {
|
||||
"model": "gpt-6-luna",
|
||||
"answers": {
|
||||
"damaged": {"type": "noul", "noul": 0.95},
|
||||
"action": {
|
||||
"type": "choice",
|
||||
"choice": "true",
|
||||
"confidence": 0.8,
|
||||
"probabilities": {"true": 0.9, "escalate": 0.1},
|
||||
},
|
||||
"severity": {
|
||||
"type": "score",
|
||||
"score": 1.72,
|
||||
"confidence": 0.58,
|
||||
"legend": {"0": "minor", "1": "major", "2": "critical"},
|
||||
"probabilities": {"0": 0.0, "1": 0.28, "2": 0.72},
|
||||
},
|
||||
},
|
||||
"usage": {"input_tokens": 383, "output_tokens": 2},
|
||||
}
|
||||
Loading…
Add table
Reference in a new issue