92 lines
3.2 KiB
Python
92 lines
3.2 KiB
Python
from __future__ import annotations
|
||
|
||
from typing import List, Tuple
|
||
|
||
from app.models import (
|
||
DataImportAnalysisRequest,
|
||
LLMMessage,
|
||
LLMProvider,
|
||
LLMRole,
|
||
)
|
||
|
||
|
||
def resolve_provider_from_model(llm_model: str) -> Tuple[LLMProvider, str]:
|
||
"""Resolve provider based on the llm_model string.
|
||
|
||
The llm_model may be provided as 'provider:model' or 'provider/model'.
|
||
If no provider prefix is present, try an educated guess from common model name patterns.
|
||
"""
|
||
|
||
normalized = llm_model.strip()
|
||
provider_hint: str | None = None
|
||
model_name = normalized
|
||
|
||
for delimiter in (":", "/", "|"):
|
||
if delimiter in normalized:
|
||
provider_hint, model_name = normalized.split(delimiter, 1)
|
||
provider_hint = provider_hint.strip().lower()
|
||
model_name = model_name.strip()
|
||
break
|
||
|
||
provider_map = {provider.value: provider for provider in LLMProvider}
|
||
|
||
if provider_hint:
|
||
if provider_hint not in provider_map:
|
||
raise ValueError(
|
||
f"Unsupported provider '{provider_hint}'. Expected one of: {', '.join(provider_map.keys())}."
|
||
)
|
||
return provider_map[provider_hint], model_name
|
||
|
||
return _guess_provider_from_model(model_name), model_name
|
||
|
||
|
||
def _guess_provider_from_model(model_name: str) -> LLMProvider:
|
||
lowered = model_name.lower()
|
||
|
||
if lowered.startswith(("gpt", "o1", "text-", "dall-e", "whisper")):
|
||
return LLMProvider.OPENAI
|
||
if lowered.startswith(("claude", "anthropic")):
|
||
return LLMProvider.ANTHROPIC
|
||
if lowered.startswith(("gemini", "models/gemini")):
|
||
return LLMProvider.GEMINI
|
||
if lowered.startswith("qwen"):
|
||
return LLMProvider.QWEN
|
||
if lowered.startswith("deepseek"):
|
||
return LLMProvider.DEEPSEEK
|
||
if lowered.startswith(("openrouter", "router-")):
|
||
return LLMProvider.OPENROUTER
|
||
|
||
supported = ", ".join(provider.value for provider in LLMProvider)
|
||
raise ValueError(
|
||
f"Unable to infer provider from model '{model_name}'. "
|
||
f"Please prefix with 'provider:model'. Supported providers: {supported}."
|
||
)
|
||
|
||
|
||
def build_import_messages(
|
||
request: DataImportAnalysisRequest,
|
||
) -> List[LLMMessage]:
|
||
"""Create system and user messages for the import analysis prompt."""
|
||
headers_formatted = "\n".join(f"- {header}" for header in request.table_headers)
|
||
|
||
system_prompt = (
|
||
"你是一名数据导入识别助手。请根据给定的表头和示例数据，判断字段含义、"
|
||
"典型数据类型以及潜在的数据质量问题。最终请返回一个结构化的JSON。\n"
|
||
"JSON结构需包含: field_summaries (数组, 每项含 header、meaning、data_type、quality_notes), "
|
||
"detected_issues (字符串数组)，以及 overall_suggestion (字符串)。"
|
||
)
|
||
|
||
user_prompt = (
|
||
f"导入记录ID: {request.import_record_id}\n\n"
|
||
"表头信息:\n"
|
||
f"{headers_formatted}\n\n"
|
||
"示例数据:\n"
|
||
f"{request.example_data}\n\n"
|
||
"请仔细分析示例数据与表头之间的对应关系，并返回符合上述JSON结构的内容。"
|
||
)
|
||
|
||
return [
|
||
LLMMessage(role=LLMRole.SYSTEM, content=system_prompt),
|
||
LLMMessage(role=LLMRole.USER, content=user_prompt),
|
||
]
|