396 lines
13 KiB
Python
396 lines
13 KiB
Python
"""LLM 信息提取
|
||
|
||
使用 LLM 从 PDF 文本/图片、支付截图中提取结构化数据。
|
||
|
||
## 功能模块
|
||
|
||
- **统一文档提取**:使用一套提示词,LLM 自行判断文档类型(发票/支付记录/出差事前申请单等),支持 JSON 格式输出。
|
||
- **差旅信息提取**:综合多张发票、支付记录和匹配结果,提取出差事由、地点、时间等差旅相关信息。
|
||
- **缓存管理**:支持从 `.invoice_cache/` 目录加载已提取的结构化数据和匹配结果,避免重复处理。
|
||
|
||
## 对外接口
|
||
|
||
- `extract_document(file_path) -> dict` — 统一入口:从任意图片/PDF 提取信息
|
||
- `extract_travel_info(source_dir) -> dict` — 综合发票和匹配结果提取差旅信息
|
||
- `load_cache(source_dir) -> dict` — 加载缓存的结构化数据
|
||
- `load_match_result(source_dir) -> dict` — 加载发票与支付记录的匹配结果
|
||
"""
|
||
|
||
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_invoice_system_prompt,
|
||
build_normal_info_system_prompt,
|
||
build_travel_info_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=65535,
|
||
request_timeout=600.0,
|
||
is_chat_model=True,
|
||
)
|
||
|
||
|
||
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))
|
||
|
||
|
||
# ------------------------------------------------------------------
|
||
# 统一文档提取(多模态,直接传图片给 LLM)
|
||
# ------------------------------------------------------------------
|
||
|
||
|
||
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 | None = None,
|
||
image_b64s: list[str] | None = None,
|
||
blocks: list[Any] | None = None,
|
||
reasoning_effort: str = "none",
|
||
) -> str:
|
||
"""发送多模态请求到 LLM。
|
||
|
||
Args:
|
||
system_prompt: 系统提示词。
|
||
text: 用户文本(与 image_b64s 配合使用,文本在前、图片在后)。
|
||
image_b64s: base64 编码的图片列表。
|
||
blocks: 预构建的内容块列表(TextBlock/ImageBlock),传入时忽略 text 和 image_b64s。
|
||
|
||
Returns:
|
||
LLM 响应文本。
|
||
"""
|
||
from llama_index.core.base.llms.types import ImageBlock, TextBlock
|
||
from llama_index.core.llms import ChatMessage
|
||
|
||
from ..config import get_llm_config
|
||
|
||
if blocks is not None:
|
||
final_blocks = blocks
|
||
else:
|
||
text = text or ""
|
||
image_b64s = image_b64s or []
|
||
final_blocks = [TextBlock(text=text)]
|
||
for img_b64 in image_b64s:
|
||
final_blocks.append(
|
||
ImageBlock(
|
||
url=f"data:image/jpeg;base64,{img_b64}",
|
||
detail="high",
|
||
)
|
||
)
|
||
|
||
messages = [
|
||
ChatMessage(role="system", content=system_prompt),
|
||
ChatMessage(role="user", blocks=final_blocks),
|
||
]
|
||
|
||
llm_config = get_llm_config()
|
||
llm = _create_llm()
|
||
log.info(
|
||
"开始请求 LLM 多模态 (model=%s, base=%s, blocks=%d)",
|
||
llm_config["model"],
|
||
llm_config["api_base"],
|
||
len(final_blocks),
|
||
)
|
||
|
||
try:
|
||
parts = []
|
||
for resp in llm.stream_chat(
|
||
messages,
|
||
temperature=0.1,
|
||
extra_body={"reasoning_effort": reasoning_effort},
|
||
):
|
||
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_document(file_path: Path) -> dict[str, Any]:
|
||
"""统一文档提取入口:从任意图片/PDF 中提取结构化信息。
|
||
|
||
LLM 会根据统一提示词自行判断文档类型(发票/支付记录/出差事前申请单等)。
|
||
|
||
Args:
|
||
file_path: 文件路径(支持 PDF 和图片格式)。
|
||
|
||
Returns:
|
||
包含提取字段的字典。
|
||
"""
|
||
from .pdf import render_pdf_to_images
|
||
|
||
system_prompt = build_invoice_system_prompt()
|
||
user_text = f"请分析以下财务文档并提取信息:\n\n文件名: {file_path.name}"
|
||
|
||
# PDF 先渲染为图片
|
||
suffix = file_path.suffix.lower()
|
||
if suffix == ".pdf":
|
||
image_b64s = render_pdf_to_images(file_path)
|
||
else:
|
||
image_b64s = [_image_to_base64(file_path)]
|
||
|
||
if not image_b64s:
|
||
log.warning(f"文件渲染为空: {file_path.name}")
|
||
return {}
|
||
|
||
try:
|
||
response = _llm_query_multimodal(system_prompt, user_text, image_b64s)
|
||
result = _parse_json_response(response)
|
||
log.info("LLM 文档提取成功: %s", file_path.name)
|
||
return result
|
||
except Exception as e:
|
||
log.error("LLM 文档提取失败: %s (%s)", file_path.name, e)
|
||
raise
|
||
|
||
|
||
# ------------------------------------------------------------------
|
||
# 差旅信息提取
|
||
# ------------------------------------------------------------------
|
||
|
||
CACHE_DIR_NAME = ".invoice_cache"
|
||
|
||
|
||
def load_cache(source_dir: Path) -> dict[str, Any]:
|
||
"""从 JSON 缓存目录加载结构化数据,构建 source filename -> 缓存数据的映射。
|
||
|
||
Args:
|
||
source_dir: 源文件目录(包含 .invoice_cache 子目录)。
|
||
|
||
Returns:
|
||
{source_filename: extracted_data} 字典。
|
||
额外包含 "travel_info" 键(如果 travel_info.json 存在)。
|
||
"""
|
||
cache_map: dict[str, Any] = {}
|
||
cache_dir = source_dir / CACHE_DIR_NAME
|
||
if not cache_dir.exists():
|
||
return cache_map
|
||
|
||
for json_path in sorted(cache_dir.glob("*.json")):
|
||
try:
|
||
with open(json_path, encoding="utf-8") as f:
|
||
cache_data = json.load(f)
|
||
|
||
# travel_info.json / normal_info.json 结构不同,直接存储
|
||
if json_path.name in ("travel_info.json", "normal_info.json"):
|
||
cache_map[json_path.name.replace(".json", "")] = cache_data
|
||
continue
|
||
|
||
extracted = cache_data.get("extracted_data", {})
|
||
src_file = extracted.get("_source_file", "")
|
||
if src_file:
|
||
cache_map[src_file] = extracted
|
||
except Exception as e:
|
||
log.warning(f"读取缓存失败 {json_path.name}: {e}")
|
||
|
||
return cache_map
|
||
|
||
|
||
def load_match_result(source_dir: Path) -> dict[str, list[dict[str, Any]]]:
|
||
"""从 JSON 缓存目录加载发票与支付记录的匹配结果。
|
||
|
||
Args:
|
||
source_dir: 源文件目录(包含 .invoice_cache 子目录)。
|
||
|
||
Returns:
|
||
{支付记录源文件 (含金额): [发票信息列表]} 字典。
|
||
每个发票信息包含 file, type, amount 字段。
|
||
"""
|
||
cache_dir = source_dir / CACHE_DIR_NAME
|
||
match_path = cache_dir / "match_result.json"
|
||
if not match_path.exists():
|
||
return {}
|
||
|
||
try:
|
||
with open(match_path, encoding="utf-8") as f:
|
||
result: dict[str, list[dict[str, Any]]] = json.load(f)
|
||
return result
|
||
except Exception as e:
|
||
log.warning(f"读取匹配结果缓存失败: {e}")
|
||
return {}
|
||
|
||
|
||
def extract_travel_info(
|
||
source_dir: Path | None = None,
|
||
) -> dict[str, Any]:
|
||
"""根据差旅发票(bot 格式),让 LLM 提取出差相关信息。
|
||
|
||
仅支持从 JSON 缓存加载数据。
|
||
|
||
bot 格式的发票包含以下字段:
|
||
- 发票类型, invoice_no, invoice_date, item_name, spec_model
|
||
- total_amount, seller_name, person_name, person_id
|
||
- card_date, card_no, card_amount, remark
|
||
|
||
Args:
|
||
source_dir: 源文件目录(必填,包含 .invoice_cache 子目录)。
|
||
|
||
Returns:
|
||
包含出差事由、地点、交通工具、时间、住宿信息等字段的字典。
|
||
"""
|
||
# 仅从 JSON 缓存加载结构化数据
|
||
if not source_dir:
|
||
log.warning("未提供 source_dir,无法加载缓存数据")
|
||
return {}
|
||
|
||
system_prompt = build_travel_info_system_prompt()
|
||
|
||
# 构建 source filename -> 缓存数据的映射
|
||
cache_map = load_cache(source_dir)
|
||
|
||
# 加载发票与支付记录的匹配结果
|
||
match_result = load_match_result(source_dir)
|
||
|
||
# 拼接纯文本消息
|
||
parts = [
|
||
"以下是本次报销的所有源文件及其提取出的结构化数据。"
|
||
"每个源文件的数据来自 OCR 识别和发票信息提取,已按文件名分组展示。"
|
||
]
|
||
|
||
# 如果有匹配结果,作为额外上下文提供
|
||
if match_result:
|
||
parts.append(
|
||
"【发票与支付记录匹配结果】"
|
||
"以下数据已将发票信息与对应的支付记录进行关联匹配,"
|
||
"用于判断每笔支付对应的发票和商户信息。\n" + json.dumps(match_result, ensure_ascii=False, indent=2)
|
||
)
|
||
|
||
# 按源文件名提供结构化数据
|
||
for filename, extracted in cache_map.items():
|
||
parts.append(
|
||
f"【源文件: {filename}】"
|
||
"以下为从该文件提取的结构化发票/支付/申请单数据。\n" + json.dumps(extracted, ensure_ascii=False, indent=2)
|
||
)
|
||
|
||
parts.append("\n=== 请返回 JSON 格式结果 ===")
|
||
user_message = "\n".join(parts)
|
||
log.info(f"user_message: {user_message}")
|
||
try:
|
||
response = _llm_query_multimodal(
|
||
system_prompt=system_prompt,
|
||
text=user_message,
|
||
reasoning_effort="low",
|
||
)
|
||
result = _parse_json_response(response)
|
||
log.info("LLM 差旅信息提取成功")
|
||
return result
|
||
except Exception as e:
|
||
log.error("LLM 差旅信息提取失败: %s", e)
|
||
raise
|
||
|
||
|
||
# ------------------------------------------------------------------
|
||
# 普通发票信息提取
|
||
# ------------------------------------------------------------------
|
||
|
||
|
||
def extract_normal_info(
|
||
source_dir: Path | None = None,
|
||
) -> dict[str, Any]:
|
||
"""根据普通发票(非差旅),让 LLM 提取报销相关信息。
|
||
|
||
仅支持从 JSON 缓存加载数据。
|
||
|
||
Args:
|
||
source_dir: 源文件目录(必填,包含 .invoice_cache 子目录)。
|
||
|
||
Returns:
|
||
包含报销说明、发票总数、总金额、支付方式、附件清单等字段的字典。
|
||
"""
|
||
if not source_dir:
|
||
log.warning("未提供 source_dir,无法加载缓存数据")
|
||
return {}
|
||
|
||
system_prompt = build_normal_info_system_prompt()
|
||
|
||
# 构建 source filename -> 缓存数据的映射
|
||
cache_map = load_cache(source_dir)
|
||
|
||
# 加载发票与支付记录的匹配结果
|
||
match_result = load_match_result(source_dir)
|
||
|
||
# 拼接纯文本消息
|
||
parts = [
|
||
"以下是本次报销的所有源文件及其提取出的结构化数据。"
|
||
"每个源文件的数据来自 OCR 识别和发票信息提取,已按文件名分组展示。"
|
||
]
|
||
|
||
# 如果有匹配结果,作为额外上下文提供
|
||
if match_result:
|
||
parts.append(
|
||
"【发票与支付记录匹配结果】"
|
||
"以下数据已将发票信息与对应的支付记录进行关联匹配,"
|
||
"用于判断每笔支付对应的发票和商户信息。\n" + json.dumps(match_result, ensure_ascii=False, indent=2)
|
||
)
|
||
|
||
# 按源文件名提供结构化数据
|
||
for filename, extracted in cache_map.items():
|
||
parts.append(
|
||
f"【源文件: {filename}】"
|
||
"以下为从该文件提取的结构化发票/支付/申请单数据。\n" + json.dumps(extracted, ensure_ascii=False, indent=2)
|
||
)
|
||
|
||
parts.append("\n=== 请返回 JSON 格式结果 ===")
|
||
user_message = "\n".join(parts)
|
||
log.info(f"user_message: {user_message}")
|
||
try:
|
||
response = _llm_query_multimodal(
|
||
system_prompt=system_prompt,
|
||
text=user_message,
|
||
reasoning_effort="low",
|
||
)
|
||
result = _parse_json_response(response)
|
||
log.info("LLM 普通发票信息提取成功")
|
||
return result
|
||
except Exception as e:
|
||
log.error("LLM 普通发票信息提取失败: %s", e)
|
||
raise
|