feat(regulatory-info-package): 实现材料包生成工作流

review_agent/regulatory_info_package/services/field_extract.py

@@ -0,0 +1,135 @@
+from __future__ import annotations
+
+import json
+import re
+import time
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from typing import Callable
+
+from review_agent.llm import generate_completion
+from review_agent.regulatory_info_package.schemas import InstructionExtractResult
+
+
+FIELD_PATTERNS = {
+    "product_name": ("产品名称", r"产品名称[:：\s]*([^\n\r]+)"),
+    "storage_condition": ("储存条件", r"(?:储存条件|贮存条件|保存条件)[:：\s]*([^\n\r]+)"),
+    "intended_use": ("预期用途", r"预期用途[:：\s]*([^\n\r]+)"),
+    "package_specification": ("包装规格", r"(?:包装规格|规格)[:：\s]*([^\n\r]+)"),
+    "sample_type": ("样本类型", r"样本类型[:：\s]*([^\n\r]+)"),
+    "applicable_instrument": ("适用仪器", r"适用仪器[:：\s]*([^\n\r]+)"),
+    "standard_no": ("标准号", r"((?:GB|YY|WS|T/C[A-Z0-9]*)[ /T0-9.\-—]+)"),
+}
+
+
+def extract_fields_by_rules(instruction: InstructionExtractResult) -> dict[str, dict]:
+    text = "\n".join([instruction.front_text, *instruction.paragraphs, *instruction.sections.values()])
+    results: dict[str, dict] = {}
+    for key, (label, pattern) in FIELD_PATTERNS.items():
+        section_value = _value_after_label_paragraph(instruction.paragraphs, label)
+        if section_value:
+            results[key] = {
+                "label": label,
+                "value": section_value,
+                "evidence": f"【{label}】\n{section_value}",
+                "confidence": 0.82,
+                "source": "rule",
+            }
+            continue
+        match = re.search(pattern, text, flags=re.IGNORECASE)
+        if match:
+            value = _clean_value(match.group(1))
+            if value:
+                results[key] = {
+                    "label": label,
+                    "value": value,
+                    "evidence": match.group(0)[:240],
+                    "confidence": 0.75,
+                    "source": "rule",
+                }
+    return results
+
+
+def extract_fields_with_llm(instruction: InstructionExtractResult) -> dict[str, dict]:
+    prompt = (
+        "请从体外诊断试剂产品说明书中抽取字段，输出 JSON 对象，字段包括 "
+        "product_name、storage_condition、intended_use、package_specification、sample_type、applicable_instrument、standard_no。"
+        "每个字段值为 {label,value,evidence,confidence}。\n\n"
+        + instruction.front_text[:6000]
+    )
+    raw = generate_completion([{"role": "user", "content": prompt}], temperature=0.0)
+    payload = _parse_json_object(raw)
+    return {key: value for key, value in payload.items() if isinstance(value, dict)}
+
+
+def run_llm_extract_with_retry(
+    instruction: InstructionExtractResult,
+    *,
+    llm_extract_func: Callable[[InstructionExtractResult], dict[str, dict]] | None = None,
+    sleep_func: Callable[[float], None] = time.sleep,
+) -> dict[str, dict]:
+    func = llm_extract_func or extract_fields_with_llm
+    last_exc: Exception | None = None
+    for delay in [0, 1, 2]:
+        if delay:
+            sleep_func(delay)
+        try:
+            return func(instruction)
+        except Exception as exc:
+            last_exc = exc
+    if last_exc:
+        raise last_exc
+    return {}
+
+
+def run_parallel_extract(
+    instruction: InstructionExtractResult,
+    *,
+    llm_extract_func: Callable[[InstructionExtractResult], dict[str, dict]] | None = None,
+) -> dict:
+    payload = {"regex_results": {}, "llm_results": {}, "llm_error": ""}
+    with ThreadPoolExecutor(max_workers=2) as executor:
+        rule_future = executor.submit(extract_fields_by_rules, instruction)
+        llm_future = executor.submit(run_llm_extract_with_retry, instruction, llm_extract_func=llm_extract_func)
+        payload["regex_results"] = rule_future.result()
+        try:
+            payload["llm_results"] = llm_future.result()
+        except Exception as exc:
+            payload["llm_error"] = str(exc)
+    return payload
+
+
+def save_field_extract_result(path: str | Path, payload: dict) -> Path:
+    target = Path(path)
+    target.parent.mkdir(parents=True, exist_ok=True)
+    target.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+    return target
+
+
+def _clean_value(value: str) -> str:
+    cleaned = value.strip()
+    if cleaned in {"】", "】】", "】:"}:
+        return ""
+    return re.split(r"[。；;]", cleaned)[0].strip()
+
+
+def _value_after_label_paragraph(paragraphs: list[str], label: str) -> str:
+    bracketed = {f"【{label}】", f"[{label}]", label}
+    for index, text in enumerate(paragraphs):
+        stripped = text.strip()
+        if stripped in bracketed and index + 1 < len(paragraphs):
+            return _clean_value(paragraphs[index + 1])
+    return ""
+
+
+def _parse_json_object(raw: str) -> dict:
+    text = (raw or "").strip()
+    if text.startswith("```"):
+        text = text.strip("`").strip()
+        if text.lower().startswith("json"):
+            text = text[4:].strip()
+    start = text.find("{")
+    end = text.rfind("}")
+    if start == -1 or end == -1:
+        return {}
+    return json.loads(text[start : end + 1])