重构项目为LLM 驱动

This commit is contained in:
wandering
2026-06-09 16:25:20 +08:00
parent 0074975591
commit 98e3d21c83
52 changed files with 7405 additions and 1485 deletions

210
src/doc/llm_extractor.py Normal file
View File

@@ -0,0 +1,210 @@
"""
LLM 信息提取
使用 LLM 从 PDF 文本中提取结构化数据,以及从支付截图中提取刷卡信息。
支持 JSON 格式输出,字段与 CSV_COLUMNS 对齐。
对外接口:
extract_invoice_from_text(text, file_name) -> dict 从 PDF 文本提取发票信息
extract_card_info_from_image(image_path) -> dict 从支付截图提取刷卡信息
"""
from __future__ import annotations
import base64
import json
from pathlib import Path
from typing import Any, cast
from .. import get_logger
from .prompt import build_card_info_system_prompt, build_invoice_system_prompt
log = get_logger("llm_extractor")
def _create_llm() -> Any:
"""根据配置文件创建 LLM 实例。"""
try:
from llama_index.llms.openai_like import OpenAILike
except ImportError:
log.error("缺少 llama-index-llms-openai-like请执行: uv pip install llama-index-llms-openai-like")
raise
from ..config import get_llm_config
llm_config = get_llm_config()
return OpenAILike(
model=llm_config["model"],
api_base=llm_config["api_base"],
api_key=llm_config.get("api_key", "lm-studio"),
temperature=0.1,
max_tokens=8192,
request_timeout=600.0,
is_chat_model=True,
)
def _llm_query(system_prompt: str, user_content: str, max_tokens: int = 4096) -> str:
"""发送请求到 LLM 并返回完整响应文本。"""
from llama_index.core.llms import ChatMessage
from ..config import get_llm_config
messages = [
ChatMessage(role="system", content=system_prompt),
ChatMessage(role="user", content=user_content),
]
llm_config = get_llm_config()
llm = _create_llm()
log.info("开始请求 LLM (model=%s, base=%s)", llm_config["model"], llm_config["api_base"])
try:
parts = []
for resp in llm.stream_chat(
messages,
temperature=0.1,
max_tokens=max_tokens,
extra_body={"reasoning_effort": "none"},
):
delta = resp.delta
if delta:
parts.append(delta)
text = "".join(parts)
log.info("LLM 请求完成,响应总长度: %d 字符", len(text))
log.info("LLM 响应: %s", text)
return text
except Exception as e:
log.error("LLM 请求失败: %s", e)
raise
def _parse_json_response(text: str) -> dict[str, Any]:
"""从 LLM 响应中提取 JSON处理可能的 Markdown 包裹。"""
text = text.strip()
# 处理 ```json ... ``` 包裹
if "```" in text:
# 提取第一个代码块
start = text.find("```") + 3
end = text.find("```", start)
if end > start:
text = text[start:end].strip()
# 去掉可能的前缀 (如 "json")
if text.lower().startswith("json"):
text = text[4:].strip()
return cast(dict[str, Any], json.loads(text))
def extract_invoice_from_text(text: str, file_name: str = "") -> dict[str, Any]:
"""从 PDF 发票文本中提取结构化数据。
Args:
text: PDF 提取的文本内容。
file_name: 原始文件名(用于日志)。
Returns:
包含所有 CSV_COLUMNS 字段的字典。
"""
system_prompt = build_invoice_system_prompt()
user_content = f"请分析以下发票文本并提取信息:\n\n文件名: {file_name}\n\n---\n\n{text}\n\n---"
try:
response = _llm_query(system_prompt, user_content, max_tokens=4096)
result = _parse_json_response(response)
log.info("LLM 发票提取成功: %s", file_name)
return result
except Exception as e:
log.error("LLM 发票提取失败: %s (%s)", file_name, e)
raise
# ------------------------------------------------------------------
# 支付截图信息提取(多模态)
# ------------------------------------------------------------------
def _image_to_base64(image_path: Path) -> str:
"""将图片文件读取为 base64 字符串。"""
with open(image_path, "rb") as f:
return base64.b64encode(f.read()).decode("utf-8")
def _llm_query_multimodal(
system_prompt: str,
text: str,
image_b64: str,
max_tokens: int = 4096,
) -> str:
"""发送多模态请求(文本 + 图片)到 LLM。"""
from llama_index.core.base.llms.types import ImageBlock, TextBlock
from llama_index.core.llms import ChatMessage
from ..config import get_llm_config
messages = [
ChatMessage(role="system", content=system_prompt),
ChatMessage(
role="user",
blocks=[
TextBlock(text=text),
ImageBlock(
url=f"data:image/jpeg;base64,{image_b64}",
detail="high",
),
],
),
]
llm_config = get_llm_config()
llm = _create_llm()
log.info(
"开始请求 LLM 多模态 (model=%s, base=%s)",
llm_config["model"],
llm_config["api_base"],
)
try:
parts = []
for resp in llm.stream_chat(
messages,
temperature=0.1,
max_tokens=max_tokens,
extra_body={"reasoning_effort": "none"},
):
delta = resp.delta
if delta:
parts.append(delta)
text = "".join(parts)
log.info("LLM 多模态请求完成,响应总长度: %d 字符", len(text))
log.info("LLM 多模态响应: %s", text)
return text
except Exception as e:
log.error("LLM 多模态请求失败: %s", e)
raise
def extract_card_info_from_image(image_path: Path) -> dict[str, Any]:
"""从支付截图中提取刷卡信息。
Args:
image_path: 支付截图图片路径。
Returns:
包含刷卡日期、刷卡金额、公务卡号的字典。
"""
system_prompt = build_card_info_system_prompt()
user_text = f"请分析以下支付截图并提取信息:\n\n文件名: {image_path.name}"
image_b64 = _image_to_base64(image_path)
try:
response = _llm_query_multimodal(system_prompt, user_text, image_b64, max_tokens=4096)
result = _parse_json_response(response)
log.info("LLM 支付截图提取成功: %s", image_path.name)
return result
except Exception as e:
log.error("LLM 支付截图提取失败: %s (%s)", image_path.name, e)
raise