Skip to content

ccran / lufa-contract

Go to a project
Commit 1152e468 by ccran
Options

feat: add skills;

parent 37f636e2
Showing with 992 additions and 826 deletions
.gitignore
......@@ -14,7 +14,7 @@
!**/*.xls
!workflow/**
!demo/**
!skills/**
!README.md
# Keep this file tracked
......
core/config.py
......@@ -50,10 +50,10 @@ MAX_WORKERS = 10
FILE_SUFFIX = "-审核批注"
## 关键参数**
use_non_fastgpt_llm = False
use_non_fastgpt_llm = True
use_lufa = False
use_jp_machine = True
debug_mode = True
debug_mode = False
## 关键参数**
max_model_len = 131072
......
data/rules.xlsx
No preview for this file type
skills/common-skill/scripts/common_tool.py deleted 100644 → 0
#!/usr/bin/env python3
"""Compatibility dispatcher for split common CLIs."""
from __future__ import annotations
import argparse, subprocess, sys
from pathlib import Path
TEXT_COMMANDS = {"random-str", "format-now", "extract-url-file", "adjust-single-chunk-size", "extract-json", "remove-duplicates-by-key", "extract-drop-json-part", "group-chunk-by-len"}
FILE_COMMANDS = {"url-replace-fastgpt", "download", "upload", "fastgpt-chat"}
def script(name: str) -> str:
return str(Path(__file__).resolve().with_name(name))
def main(argv: list[str] | None = None) -> int:
args = list(sys.argv[1:] if argv is None else argv)
if not args or args[0] in {"-h", "--help"}:
p = argparse.ArgumentParser(description="Compatibility dispatcher for json_text_tool.py and file_chat_tool.py")
p.add_argument("command", nargs="?", choices=sorted(TEXT_COMMANDS | FILE_COMMANDS)); p.print_help()
print("\nText/JSON commands:", ", ".join(sorted(TEXT_COMMANDS)))
print("File/chat commands:", ", ".join(sorted(FILE_COMMANDS)))
return 0
if args[0] in TEXT_COMMANDS:
target = script("json_text_tool.py")
elif args[0] in FILE_COMMANDS:
target = script("file_chat_tool.py")
else:
print(f"unknown command: {args[0]}", file=sys.stderr); return 2
return subprocess.call([sys.executable, target, *args])
if __name__ == "__main__":
raise SystemExit(main())
skills/common-skill/scripts/file_chat_tool.py deleted 100644 → 0
#!/usr/bin/env python3
"""Standalone file and FastGPT chat utility CLI."""
from __future__ import annotations
import argparse, json, mimetypes, random, re, string, sys, time, urllib.error, urllib.parse, urllib.request
from pathlib import Path
from typing import Any
FASTGPT = "http://172.21.107.45:3030"
BACKEND = "http://172.21.107.45:1122"
OUTER = "https://172.21.107.45:48080"
def rand(n: int = 8) -> str:
return "".join(random.choice(string.ascii_lowercase) for _ in range(n))
def url_replace_fastgpt(origin: str, base_fastgpt_url: str = FASTGPT) -> str:
return origin if origin.startswith(("http:", "https:")) else base_fastgpt_url + origin
def basename(name: str) -> str:
return Path(urllib.parse.unquote(name.strip().strip('"')).replace("\\", "/")).name or "downloaded_file"
def resolve_name(url: str, headers: dict[str, str]) -> str:
cd = headers.get("content-disposition", "") or headers.get("Content-Disposition", "")
for pat in [r"filename\*=(?:UTF-8''|utf-8'')?([^;]+)", r'filename="?([^";]+)"?']:
m = re.search(pat, cd)
if m:
return basename(m.group(1))
return basename(urllib.parse.urlparse(url).path)
def download_file(url: str, path: str, input_url_to_inner: bool = True, base_fastgpt_url: str = FASTGPT, base_backend_url: str = BACKEND, outer_backend_url: str = OUTER) -> str | None:
if input_url_to_inner and not url.startswith(("http:", "https:")):
url = base_fastgpt_url + url
if input_url_to_inner:
url = url.replace(outer_backend_url, base_backend_url)
try:
with urllib.request.urlopen(urllib.request.Request(url, method="GET"), timeout=120) as resp:
target = Path(path)
if target.exists() and target.is_dir():
target = target / resolve_name(url, dict(resp.headers))
target.parent.mkdir(parents=True, exist_ok=True); target.write_bytes(resp.read()); return str(target)
except urllib.error.HTTPError as exc:
print(f"{url}文件下载失败. HTTP Status Code: {exc.code}", file=sys.stderr); return None
def post_json(url: str, data: dict[str, Any], headers: dict[str, str] | None = None, timeout: int = 120) -> str:
req = urllib.request.Request(url, data=json.dumps(data, ensure_ascii=False).encode(), headers={"Content-Type": "application/json", **(headers or {})}, method="POST")
with urllib.request.urlopen(req, timeout=timeout) as resp:
return resp.read().decode("utf-8", errors="replace")
def multipart(path: str) -> tuple[bytes, str]:
p = Path(path); boundary = f"----common-tool-{int(time.time() * 1000)}-{rand()}"
ctype = mimetypes.guess_type(p.name)[0] or "application/octet-stream"
body = bytearray(f'--{boundary}\r\nContent-Disposition: form-data; name="file"; filename="{p.name}"\r\nContent-Type: {ctype}\r\n\r\n'.encode())
body.extend(p.read_bytes()); body.extend(f"\r\n--{boundary}--\r\n".encode()); return bytes(body), boundary
def upload_file(path: str, base_backend_url: str = BACKEND, username: str = "admin", password: str = "admin@jpai.com") -> str:
token = (json.loads(post_json(f"{base_backend_url}/admin-api/system/auth/login", {"username": username, "password": password})).get("data") or {}).get("accessToken")
if not token:
raise RuntimeError("后端登录异常")
body, boundary = multipart(path)
req = urllib.request.Request(f"{base_backend_url}/admin-api/infra/file/upload", data=body, headers={"Content-Type": f"multipart/form-data; boundary={boundary}", "Authorization": token}, method="POST")
with urllib.request.urlopen(req, timeout=120) as resp:
text = resp.read().decode("utf-8", errors="replace")
res = json.loads(text).get("data")
if not res:
raise RuntimeError(f"上传{path}失败 Response text: {text}")
return res
def fastgpt_openai_chat(url: str, token: str, model: str, chat_id: str, file_url: str, text: str, stream: bool = True) -> str:
data = {"chatId": chat_id, "messages": [{"role": "user", "content": [{"type": "file_url", "name": "文件", "url": file_url}, {"type": "text", "text": text}]}], "model": model, "stream": stream}
req = urllib.request.Request(url, data=json.dumps(data, ensure_ascii=False).encode(), headers={"Content-Type": "application/json", "Authorization": f"Bearer {token}"}, method="POST")
with urllib.request.urlopen(req, timeout=60000) as resp:
if not stream:
rsp = json.loads(resp.read().decode("utf-8", errors="replace")); return rsp.get("choices", [{}])[0].get("message", {}).get("content", "")
out = ""
for raw in resp:
line = raw.decode("utf-8", errors="replace").strip()
if not line or line == "data: [DONE]": continue
try:
out += json.loads(line[6:] if line.startswith("data: ") else line).get("choices", [{}])[0].get("delta", {}).get("content", "")
except Exception:
pass
return out
def main() -> int:
p = argparse.ArgumentParser(description="File/FastGPT utilities"); sub = p.add_subparsers(dest="cmd", required=True)
a = sub.add_parser("url-replace-fastgpt"); a.add_argument("origin"); a.add_argument("--base-fastgpt-url", default=FASTGPT)
a = sub.add_parser("download"); a.add_argument("url"); a.add_argument("path"); a.add_argument("--base-fastgpt-url", default=FASTGPT); a.add_argument("--base-backend-url", default=BACKEND); a.add_argument("--outer-backend-url", default=OUTER); a.add_argument("--no-input-url-to-inner", action="store_true")
a = sub.add_parser("upload"); a.add_argument("path"); a.add_argument("--base-backend-url", default=BACKEND); a.add_argument("--username", default="admin"); a.add_argument("--password", default="admin@jpai.com")
a = sub.add_parser("fastgpt-chat"); a.add_argument("--url", required=True); a.add_argument("--token", required=True); a.add_argument("--model", required=True); a.add_argument("--chat-id", required=True); a.add_argument("--file-url", required=True); a.add_argument("--text", required=True); a.add_argument("--no-stream", action="store_true")
x = p.parse_args()
if x.cmd == "url-replace-fastgpt": print(url_replace_fastgpt(x.origin, x.base_fastgpt_url))
elif x.cmd == "download": print(download_file(x.url, x.path, not x.no_input_url_to_inner, x.base_fastgpt_url, x.base_backend_url, x.outer_backend_url))
elif x.cmd == "upload": print(upload_file(x.path, x.base_backend_url, x.username, x.password))
elif x.cmd == "fastgpt-chat": print(fastgpt_openai_chat(x.url, x.token, x.model, x.chat_id, x.file_url, x.text, not x.no_stream))
return 0
if __name__ == "__main__":
raise SystemExit(main())
skills/common-skill/scripts/json_text_tool.py deleted 100644 → 0
#!/usr/bin/env python3
"""Standalone JSON/text utility CLI."""
from __future__ import annotations
import argparse, json, random, re, string, sys
from datetime import datetime
from pathlib import Path
from typing import Any
MIN_SIZE, MAX_SIZE, MAX_PAGE = 2000, 100000, 10
def random_str(n: int = 5) -> str:
return "".join(random.choice(string.ascii_lowercase) for _ in range(n)) if n > 26 else "".join(random.sample(string.ascii_lowercase, n))
def format_now() -> str:
return datetime.now().strftime("%Y-%m-%d %H:%M:%S")
def extract_url_file(url: str, formats: list[str]) -> str:
pat = "|".join(r"[\u4e00-\u9fa5（）()0-9\w-]+" + re.escape(f) for f in formats)
m = re.search(pat, url)
if not m:
raise RuntimeError(f"{formats} not found in url:{url}")
return m.group()
def adjust_single_chunk_size(length: int, max_page: int = MAX_PAGE, min_size: int = MIN_SIZE, max_size: int = MAX_SIZE) -> int:
return max(min_size, min(length // max_page, max_size))
def _loads(text: str) -> Any:
try:
import json_repair # type: ignore
return json_repair.loads(text, strict=False)
except ImportError:
return json.loads(text)
def extract_json(text: str) -> list[Any]:
def add(candidate: str, out: list[Any]) -> bool:
s = re.sub(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]", "", (candidate or "").strip())
if not s:
return False
try:
obj = _loads(s)
except Exception:
return False
out.extend(obj if isinstance(obj, list) else [obj]); return True
out: list[Any] = []
for m in re.findall(r"```json([\s\S]*?)```", text or "", re.DOTALL):
add(m, out)
if out or add(text or "", out):
return out
for m in re.findall(r"```([\s\S]*?)```", text or "", re.DOTALL):
if add(m, out):
return out
for m in re.findall(r"(\{[\s\S]*?\}|\[[\s\S]*?\])", text or "", re.DOTALL):
add(m, out)
return out
def remove_duplicates_by_key(items: list[dict[str, Any]], key: str) -> list[dict[str, Any]]:
out, seen = [], []
for item in sorted(items, key=lambda x: len(str(x.get(key, ""))), reverse=True):
v = str(item.get(key, ""))
if not any(v in s for s in seen):
seen.append(v); out.append(item)
return out
def group_chunk_by_len(items: list[dict[str, Any]], key: str, chunk_len: int) -> list[list[dict[str, Any]]]:
groups, current, acc = [], [], 0
for item in items:
n = len(str(item.get(key, "")))
if current and acc + n > chunk_len:
groups.append(current); current, acc = [], 0
current.append(item); acc += n
return groups + ([current] if current else [])
def read_json_arg(value: str) -> Any:
p = Path(value)
return json.loads(p.read_text(encoding="utf-8")) if p.exists() else json.loads(value)
def main() -> int:
p = argparse.ArgumentParser(description="JSON/text utilities"); sub = p.add_subparsers(dest="cmd", required=True)
a = sub.add_parser("random-str"); a.add_argument("-l", "--length", type=int, default=5)
sub.add_parser("format-now")
a = sub.add_parser("extract-url-file"); a.add_argument("url"); a.add_argument("formats", nargs="+")
a = sub.add_parser("adjust-single-chunk-size"); a.add_argument("all_text_len", type=int); a.add_argument("--max-chunk-page", type=int, default=MAX_PAGE); a.add_argument("--min-single-chunk-size", type=int, default=MIN_SIZE); a.add_argument("--max-single-chunk-size", type=int, default=MAX_SIZE)
a = sub.add_parser("extract-json"); a.add_argument("text", nargs="?")
a = sub.add_parser("remove-duplicates-by-key"); a.add_argument("json_list"); a.add_argument("key")
a = sub.add_parser("extract-drop-json-part"); a.add_argument("text", nargs="?")
a = sub.add_parser("group-chunk-by-len"); a.add_argument("json_list"); a.add_argument("key"); a.add_argument("chunk_len", type=int)
x = p.parse_args()
if x.cmd == "random-str": print(random_str(x.length))
elif x.cmd == "format-now": print(format_now())
elif x.cmd == "extract-url-file": print(extract_url_file(x.url, x.formats))
elif x.cmd == "adjust-single-chunk-size": print(adjust_single_chunk_size(x.all_text_len, x.max_chunk_page, x.min_single_chunk_size, x.max_single_chunk_size))
elif x.cmd == "extract-json": print(json.dumps(extract_json(x.text if x.text is not None else sys.stdin.read()), ensure_ascii=False, indent=2))
elif x.cmd == "remove-duplicates-by-key": print(json.dumps(remove_duplicates_by_key(read_json_arg(x.json_list), x.key), ensure_ascii=False, indent=2))
elif x.cmd == "extract-drop-json-part": print(re.sub(r"```json([\s\S]*?)```", "", x.text if x.text is not None else sys.stdin.read(), flags=re.DOTALL).strip())
elif x.cmd == "group-chunk-by-len": print(json.dumps(group_chunk_by_len(read_json_arg(x.json_list), x.key, x.chunk_len), ensure_ascii=False, indent=2))
return 0
if __name__ == "__main__":
raise SystemExit(main())
skills/contract-review-flow-skill/scripts/review_flow.py deleted 100644 → 0
#!/usr/bin/env python3
"""Standalone contract review orchestration CLI."""
from __future__ import annotations
import argparse, json, subprocess, sys
from pathlib import Path
from typing import Any
ROOT = Path(__file__).resolve().parents[3]
DOC = ROOT / "skills/doc-excel-skill/scripts/doc_tool.py"
XLS = ROOT / "skills/doc-excel-skill/scripts/excel_tool.py"
LLM = ROOT / "skills/review-llm-skill/scripts/review_llm_skill.py"
COLS = {"id": "ID", "title": "审查项", "rule": "审查规则", "level": "风险等级", "triggers": "触发词", "suggestion_template": "建议模板", "case": "案例", "summary": "摘要项"}
def sh(args: list[str], text: bool = False) -> Any:
out = subprocess.check_output([sys.executable, *args], text=True)
return out if text else json.loads(out or "null")
def dump(path: Path, data: Any) -> str:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(data, ensure_ascii=False, indent=2), encoding="utf-8")
return str(path)
def norm_rules(rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
return [{k: r.get(v, "") for k, v in COLS.items()} for r in rows if isinstance(r, dict)]
def titles(items: list[dict[str, Any]]) -> list[str]:
return [str(i.get("title") or i.get("rule_title") or "").strip() for i in items if str(i.get("title") or i.get("rule_title") or "").strip()]
def pick_rules(all_rules: list[dict[str, Any]], selected: list[str]) -> list[dict[str, Any]]:
if not selected:
return all_rules
selected_set = set(selected)
return [r for r in all_rules if r.get("title") in selected_set]
def llm(tool: str, **kw: Any) -> dict[str, Any]:
args = [str(LLM), tool]
for key, value in kw.items():
flag = "--" + key.replace("_", "-")
if isinstance(value, (dict, list)):
args += [flag, json.dumps(value, ensure_ascii=False)]
elif value is not None:
args += [flag, str(value)]
return sh(args)
def route_segment(text: str, rules: list[dict[str, Any]], party_role: str, mode: str) -> tuple[list[str], list[str], list[dict[str, Any]]]:
if mode == "none":
return titles(rules), sorted({r.get("summary", "") for r in rules if r.get("summary")}), rules
res = llm("router", segment_text=text, rules=rules, party_role=party_role)
items = res.get("selected_items") or res.get("routed_rules") or []
sel_titles = titles(items)
routed = pick_rules(rules, sel_titles)
summaries = sorted({r.get("summary", "") for r in routed if r.get("summary")})
return titles(routed), summaries, routed
def reflect_findings(rules: list[dict[str, Any]], facts: list[dict[str, Any]], findings: list[dict[str, Any]], party_role: str) -> list[dict[str, Any]]:
out: list[dict[str, Any]] = []
for rule in rules:
name = rule.get("title", "")
scoped = [f for f in findings if f.get("rule_title") == name]
if not scoped:
continue
summary = rule.get("summary", "")
fact_scope = [{summary: f.get(summary)} for f in facts if summary and isinstance(f, dict) and summary in f]
res = llm("reflect", rule=rule, findings=scoped, facts=fact_scope, party_role=party_role)
out.extend(res.get("final_findings") or res.get("findings") or [])
return out
def merge_by_segment(findings: list[dict[str, Any]]) -> list[dict[str, Any]]:
out: list[dict[str, Any]] = []
segs = sorted({int(f.get("segment_id", 0) or 0) for f in findings})
for seg in segs:
items = [f for f in findings if int(f.get("segment_id", 0) or 0) == seg and str(f.get("result", "")).strip() == "不合格"]
if not items:
continue
res = llm("merger", payload=items)
merged = res.get("findings") or []
for item in merged:
item.setdefault("segment_id", seg)
out.extend(merged)
return out
def run(file: Path, rules_path: Path, ruleset: str, out_dir: Path, party_role: str, route_by: str, reflect: bool, merge_mode: str, max_chunks: int, dry_run: bool) -> dict[str, Any]:
out_dir.mkdir(parents=True, exist_ok=True)
info = sh([str(DOC), "doc-load", str(file)])
rows = sh([str(XLS), "load-excel", str(rules_path), "--sheet-name", ruleset])
rules, chunk_ids = norm_rules(rows), info.get("chunk_ids", [])
if max_chunks:
chunk_ids = chunk_ids[:max_chunks]
memory = {"file": str(file), "ruleset": ruleset, "segment_ids": [i + 1 for i in chunk_ids], "rule_titles": titles(rules), "summary_names": sorted({r.get("summary", "") for r in rules if r.get("summary")}), "facts": [], "merge_facts": [], "findings": {"review": [], "reflect": [], "merge": []}}
if dry_run:
dump(out_dir / "memory.json", memory); return memory
for cid in chunk_ids:
text = sh([str(DOC), "doc-chunk", str(file), str(cid)], text=True)
routed_titles, routed_summaries, routed_rules = route_segment(text, rules, party_role, route_by)
summary = llm("summary", segment_text=text, rules=routed_rules, party_role=party_role)
fact = summary.get("facts", summary)
if isinstance(fact, dict):
memory["facts"].append(fact)
review = llm("review", segment_text=text, rules=routed_rules, party_role=party_role)
for f in review.get("findings", []):
f.setdefault("segment_id", cid); memory["findings"]["review"].append(f)
memory.setdefault("routes", []).append({"segment_id": cid + 1, "routed_rule_titles": routed_titles, "routed_summary_names": routed_summaries})
if reflect:
memory["findings"]["reflect"] = reflect_findings(rules, memory["facts"], memory["findings"]["review"], party_role)
source = memory["findings"]["reflect"] or memory["findings"]["review"]
memory["findings"]["merge"] = merge_by_segment(source)
fact_res = llm("fact-merge", facts=memory["facts"], summary_names=memory["summary_names"], merge_mode=merge_mode)
memory["merge_facts"] = [fact_res.get("merge_facts", {})]
mem_path = Path(dump(out_dir / "memory.json", memory))
sh([str(XLS), "export-findings-excel", "@" + str(mem_path), str(out_dir / "review.xlsx"), "--finding-key", "merge" if memory["findings"]["merge"] else ("reflect" if memory["findings"]["reflect"] else "review")], text=True)
if file.suffix.lower() == ".docx":
sh([str(DOC), "docx-add-comments", str(file), "@" + str(mem_path), str(out_dir / "commented.docx"), "--finding-key", "merge" if memory["findings"]["merge"] else ("reflect" if memory["findings"]["reflect"] else "review")], text=True)
return memory
def main() -> int:
p = argparse.ArgumentParser(description="Contract review flow orchestrator")
p.add_argument("file"); p.add_argument("--rules", default=str(ROOT / "data/rules.xlsx")); p.add_argument("--ruleset", default="通用"); p.add_argument("--out-dir", default="outputs/review-flow")
p.add_argument("--party-role", default=""); p.add_argument("--route-by", choices=["rule", "none"], default="rule"); p.add_argument("--no-reflect", action="store_true"); p.add_argument("--merge-mode", choices=["llm", "rule"], default="rule")
p.add_argument("--max-chunks", type=int, default=0); p.add_argument("--dry-run", action="store_true")
a = p.parse_args()
run(Path(a.file), Path(a.rules), a.ruleset, Path(a.out_dir), a.party_role, a.route_by, not a.no_reflect, a.merge_mode, a.max_chunks, a.dry_run)
print("输出目录:", Path(a.out_dir).resolve())
return 0
if __name__ == "__main__":
raise SystemExit(main())
skills/doc-excel-skill/SKILL.md 0 → 100644
---
name: doc-excel-skill
description: 文档/表格工具 Skill。用于将 Word/PDF 文档解析为 txt，并读取、修改 Excel，以及把 JSON 写入 Excel sheet。
---
# 文档与 Excel Skill
## 定位
`doc-excel-skill` 负责文件解析和表格读写。它把 Word/PDF 文件转换为 UTF-8 `.txt`，把 Excel sheet 读取为结构化 JSON，也可以把 JSON 数据写入 Excel sheet。
该 Skill 不直接调用 LLM，也不实现业务审查逻辑。它提供的是稳定的文件 I/O 能力，供 `review-llm-skill` 和 `contract-review-flow-skill` 组合使用。
## 适用场景
- 将 `.docx`、`.doc`、`.wps`、`.pdf` 合同文件解析为 `.txt`。
- 读取 `data/rules.xlsx` 中的规则表。
- 列出 Excel sheets，并按条件 dict 搜索 sheet 行数据。
- 按表头 key 追加 Excel 行，或按条件 dict 更新、删除 Excel 行。
- 按列查找 Excel 单元格，或将 Excel 行映射为指定字段。
- 将 JSON 转换为 Excel 的某个 sheet。
## 工具文件
- `scripts/doc_tool.py`：基于 Spire 的 Word/PDF 转 txt CLI。
- `scripts/excel_tool.py`：Excel 读取、sheet 查询、行增删改、JSON 写入 sheet。
## 依赖说明
- Word / PDF 文本解析依赖 Spire：PDF 使用 `PdfDocument`、`PdfTextExtractOptions`、`PdfTextExtractor`；Word 使用 `Document.GetText()`。
- 如果安装了 `openpyxl`，Excel 读取和写入会优先使用它。
- 如果没有 `openpyxl`，部分 `.xlsx` 读取会退回标准库实现，但复杂写入能力会受限。
- 当前 `doc_tool.py` 只做可提取文本解析，不做 OCR、分块和批注写入。
## 主要命令
- `doc_tool.py <file> [output]`：将 Word/PDF 解析为 txt；未传 `output` 时默认输出到同名 `.txt`。
- `doc_tool.py doc-to-txt <file> [output]`：兼容旧调用形式，行为同上。
- `load-excel`：读取 Excel sheet 为 JSON。
- `list-sheets`：列出工作簿中的 sheet。
- `search_rows`：读取某个 sheet，按首行表头作为 key、每一行作为 dict；传入条件 dict 时返回所有匹配行，传 `{}` 时返回全部行。
- `append-row`：传入一个 dict，按 key 匹配表头列，并追加到指定 sheet 末尾。
- `update-rows`：传入条件 dict 和更新 dict，更新所有匹配行。
- `delete-rows`：传入条件 dict，删除所有匹配行。
- `find-value`：按某列匹配值，再返回另一列的值。
- `map-rows`：按字段映射读取 Excel 行。
- `json-to-sheet`：把 JSON 写入 Excel 的指定 sheet；dict key 会直接写入第一行作为表头。
## 输入输出
- 文档解析命令输入本地 Word/PDF 文件路径，输出 UTF-8 `.txt` 文件路径。
- Excel 读取类命令输入 `.xlsx` 路径和 sheet/列参数，输出 JSON。
- Excel 写入类命令输入 `.xlsx` 路径、sheet 名和 JSON dict，直接保存原文件。
- 行搜索、更新和删除的条件 dict 支持多个字段，所有字段都相等时才算匹配。
- `json-to-sheet` 输入 JSON 和输出 `.xlsx` 路径；如果目标 sheet 已存在，会替换该 sheet；第一行固定写表头，第二行开始写数据。
- 支持使用 `@file.json` 形式读取较大的 JSON 参数。
## 使用示例
查看帮助：
```bash
python skills/doc-excel-skill/scripts/doc_tool.py --help
python skills/doc-excel-skill/scripts/excel_tool.py --help
```
解析 Word/PDF 合同为同名 txt：
```bash
python skills/doc-excel-skill/scripts/doc_tool.py demo/example.docx
python skills/doc-excel-skill/scripts/doc_tool.py demo/example.pdf
```
解析 Word/PDF 合同到指定 txt：
```bash
python skills/doc-excel-skill/scripts/doc_tool.py demo/example.docx outputs/example.txt
python skills/doc-excel-skill/scripts/doc_tool.py doc-to-txt demo/example.pdf outputs/example.txt
```
读取规则 Excel：
```bash
python skills/doc-excel-skill/scripts/excel_tool.py load-excel data/rules.xlsx \
--sheet-name 通用
```
列出所有 sheet：
```bash
python skills/doc-excel-skill/scripts/excel_tool.py list-sheets data/rules.xlsx
```
按条件搜索指定 sheet 的行，传 `{}` 返回全部行：
```bash
python skills/doc-excel-skill/scripts/excel_tool.py search_rows data/rules.xlsx \
'{"审查项":"当事人审查"}' \
--sheet-name 通用
```
追加一行：
```bash
python skills/doc-excel-skill/scripts/excel_tool.py append-row data/rules.xlsx \
'{"审查项":"测试","风险等级":"中"}' \
--sheet-name 通用
```
更新一行：
```bash
python skills/doc-excel-skill/scripts/excel_tool.py update-rows data/rules.xlsx \
'{"审查项":"测试","风险等级":"中"}' \
'{"风险等级":"高"}' \
--sheet-name 通用
```
删除一行：
```bash
python skills/doc-excel-skill/scripts/excel_tool.py delete-rows data/rules.xlsx \
'{"审查项":"测试","风险等级":"高"}' \
--sheet-name 通用
```
将 JSON 写入 Excel sheet：
```bash
python skills/doc-excel-skill/scripts/excel_tool.py json-to-sheet \
'[{"name":"张三","amount":100},{"name":"李四","amount":200}]' \
outputs/result.xlsx \
--sheet-name 明细
```
从 JSON 文件写入 Excel sheet：
```bash
python skills/doc-excel-skill/scripts/excel_tool.py json-to-sheet \
@data.json \
outputs/result.xlsx \
--sheet-name 数据
```
## 在合同审查流程中的位置
该 Skill 主要对应文档解析和 Excel 数据读写部分。它可以在处理前把 Word/PDF 转换为 txt，也可以把 Excel sheet 和 JSON 数据在两种结构之间转换。
skills/doc-excel-skill/scripts/doc_excel_tool.py deleted 100644 → 0
#!/usr/bin/env python3
"""Compatibility dispatcher for the split doc/excel CLIs."""
from __future__ import annotations
import argparse
import subprocess
import sys
from pathlib import Path
EXCEL_COMMANDS = {
"load-excel",
"list-sheets",
"find-value",
"map-rows",
"export-findings-excel",
"export-facts-excel",
}
DOC_COMMANDS = {
"doc-load",
"doc-ocr",
"doc-chunk",
"doc-info",
"doc-adjust-chunk-size",
"doc-text",
"docx-add-comments",
"process-string",
"is-messy-text",
}
def _script(name: str) -> str:
return str(Path(__file__).resolve().with_name(name))
def main(argv: list[str] | None = None) -> int:
args = list(sys.argv[1:] if argv is None else argv)
if not args or args[0] in {"-h", "--help"}:
parser = argparse.ArgumentParser(description="Compatibility dispatcher for doc_tool.py and excel_tool.py")
parser.add_argument("command", nargs="?", choices=sorted(EXCEL_COMMANDS | DOC_COMMANDS))
parser.print_help()
print("\nExcel commands:", ", ".join(sorted(EXCEL_COMMANDS)))
print("Doc commands:", ", ".join(sorted(DOC_COMMANDS)))
return 0
command = args[0]
if command in EXCEL_COMMANDS:
target = _script("excel_tool.py")
elif command in DOC_COMMANDS:
target = _script("doc_tool.py")
else:
print(f"unknown command: {command}", file=sys.stderr)
return 2
return subprocess.call([sys.executable, target, *args])
if __name__ == "__main__":
raise SystemExit(main())
skills/http-skill/SKILL.md 0 → 100644
---
name: http-skill
description: HTTP 文件处理 Skill。用于下载远程文件和上传本地文件。上传下载优先使用此技能。
---
# HTTP 文件处理 Skill
## 定位
`http-skill` 负责合同审查流程中的网络文件搬运。它可以把接口传入的远程合同 URL 下载到本地，也可以把审查结果文件上传到后端服务。
该 Skill 使用 Python 标准库实现，不依赖 `requests`、`loguru`、`requests_toolbelt`，也不依赖仓库中的 `utils/` 或 `core/` 模块。
## 适用场景
- 从接口 URL 下载合同、PDF、Excel 或中间文件。
- 将本地生成的 Excel、docx 批注文件上传到后端文件服务。
- 在离线 CLI 流程中模拟 `main.py` 的文件下载和导出上传环节。
## 工具文件
- `scripts/http_util.py`：HTTP 文件处理 CLI。
## 主要命令
- `download`：下载 URL 到本地文件或目录。
- `upload`：上传本地文件到后端文件服务。
## 通用参数
- `--base-fastgpt-url`：FastGPT 内网基础地址，默认 `http://192.168.252.71:3030`。
- `--base-backend-url`：后端内网基础地址，默认 `http://192.168.252.71:1122`。
- `--outer-backend-url`：后端外网地址，默认 `https://218.77.58.8:48080`。
- `--username`：后端管理员用户名，仅 `upload` 使用，默认 `admin`。
- `--password`：后端管理员密码，仅 `upload` 使用，默认 `admin@jpai.com`。
## 输入输出
- `download` 输入 URL 和可选目标路径；未传目标路径时默认下载到 `scripts/http_util.py` 同级目录的 `download/` 文件夹。
- `download` 的目标路径是目录时，会自动推断文件名。
- `upload` 输入本地文件路径和后端账号配置；输出后端接口响应。
- `upload` 输入相对路径时，会优先从 `scripts/http_util.py` 同级目录的 `download/` 文件夹查找，找不到再按当前工作目录查找。
## 使用示例
查看帮助：
```bash
python skills/http-skill/scripts/http_util.py --help
```
查看子命令帮助：
```bash
python skills/http-skill/scripts/http_util.py upload --help
python skills/http-skill/scripts/http_util.py download --help
```
上传本地文件：
```bash
python skills/http-skill/scripts/http_util.py upload demo/example.pdf
```
上传本地文件，并覆盖后端地址和账号密码：
```bash
python skills/http-skill/scripts/http_util.py upload \
--base-backend-url http://192.168.252.71:48081 \
--username admin \
--password 'admin@jpai.com' \
demo/example.pdf
```
下载相对路径到默认 `download/` 目录：
```bash
python skills/http-skill/scripts/http_util.py download /api/file/example.pdf
```
下载相对路径到指定目录：
```bash
python skills/http-skill/scripts/http_util.py download \
/api/file/example.pdf \
downloads
```
下载完整 URL，并替换外网后端地址：
```bash
python skills/http-skill/scripts/http_util.py download \
--outer-backend-url https://172.21.107.45:48080 \
--base-backend-url http://172.21.107.45:1122 \
https://172.21.107.45:48080/admin-api/infra/file/get/123 \
downloads/example.pdf
```
## 在合同审查流程中的位置
该 Skill 通常位于流程入口和出口：入口负责把远程合同下载成本地文件，出口负责把审查结果上传并生成可返回给调用方的文件地址。它不解析文档、不调用 LLM，也不保存审查记忆。
skills/http-skill/scripts/http_util.py
......@@ -7,23 +7,29 @@ import argparse, json, mimetypes, random, re, string, sys, time, urllib.error, u
from pathlib import Path
from urllib.parse import unquote, urlparse
DEFAULT_OUTER_BACKEND_URL = "https://172.21.107.45:48080"
DEFAULT_BASE_FASTGPT_URL = "http://172.21.107.45:3030"
DEFAULT_BASE_BACKEND_URL = "http://172.21.107.45:1122"
# DEFAULT_OUTER_BACKEND_URL = "https://172.21.107.45:48080"
# DEFAULT_BASE_FASTGPT_URL = "http://172.21.107.45:3030"
# DEFAULT_BASE_BACKEND_URL = "http://172.21.107.45:1122"
DEFAULT_OUTER_BACKEND_URL = "https://218.77.58.8:48080"
DEFAULT_BASE_FASTGPT_URL = "http://192.168.252.71:3030"
DEFAULT_BASE_BACKEND_URL = "http://192.168.252.71:1122"
DEFAULT_BACKEND_ADMIN_USERNAME = "admin"
DEFAULT_BACKEND_ADMIN_PASSWORD = "admin@jpai.com"
SCRIPT_DIR = Path(__file__).resolve().parent
DEFAULT_DOWNLOAD_DIR = SCRIPT_DIR / "download"
base_fastgpt_url, base_backend_url, outer_backend_url = DEFAULT_BASE_FASTGPT_URL, DEFAULT_BASE_BACKEND_URL, DEFAULT_OUTER_BACKEND_URL
backend_admin_username, backend_admin_password = DEFAULT_BACKEND_ADMIN_USERNAME, DEFAULT_BACKEND_ADMIN_PASSWORD
def configure_urls(fastgpt_url: str | None = None, backend_url: str | None = None, outer_url: str | None = None) -> None:
def _configure_urls(fastgpt_url: str | None = None, backend_url: str | None = None, outer_url: str | None = None) -> None:
global base_fastgpt_url, base_backend_url, outer_backend_url
base_fastgpt_url = fastgpt_url or base_fastgpt_url
base_backend_url = backend_url or base_backend_url
outer_backend_url = outer_url or outer_backend_url
def configure_login(username: str | None = None, password: str | None = None) -> None:
def _configure_login(username: str | None = None, password: str | None = None) -> None:
global backend_admin_username, backend_admin_password
backend_admin_username = username or backend_admin_username
backend_admin_password = password or backend_admin_password
......@@ -53,7 +59,16 @@ def _multipart_body(path: str, field: str = "file") -> tuple[bytes, str]:
return bytes(body), boundary
def upload_file(path, input_url_to_inner=True, output_url_to_inner=False) -> str:
def _resolve_upload_path(path: str | Path) -> Path:
p = Path(path).expanduser()
if p.is_absolute():
return p
download_path = DEFAULT_DOWNLOAD_DIR / p
return download_path if download_path.exists() else p
def upload_file(path) -> str:
path = _resolve_upload_path(path)
login = _post_json(f"{base_backend_url}/admin-api/system/auth/login", {"username": backend_admin_username, "password": backend_admin_password})
token = (json.loads(login).get("data") or {}).get("accessToken")
if not token:
......@@ -81,14 +96,13 @@ def _resolve_name(url: str, headers) -> str:
return _basename(urlparse(url).path)
def download_file(url, path, input_url_to_inner=True):
if input_url_to_inner and not url.startswith(("http:", "https:")):
def download_file(url, path=None):
if not url.startswith(("http:", "https:")):
url = base_fastgpt_url + url
if input_url_to_inner:
url = url.replace(outer_backend_url, base_backend_url)
try:
with urllib.request.urlopen(urllib.request.Request(url, method="GET"), timeout=120) as resp:
target = Path(path)
target = Path(path).expanduser() if path else DEFAULT_DOWNLOAD_DIR / _resolve_name(url, resp.headers)
if target.exists() and target.is_dir():
target = target / _resolve_name(url, resp.headers)
target.parent.mkdir(parents=True, exist_ok=True); target.write_bytes(resp.read())
......@@ -98,37 +112,30 @@ def download_file(url, path, input_url_to_inner=True):
return None
def url_replace_fastgpt(origin: str):
return origin if origin.startswith(("http:", "https:")) else base_fastgpt_url + origin
def add_url_args(p: argparse.ArgumentParser) -> None:
def _add_url_args(p: argparse.ArgumentParser) -> None:
p.add_argument("--base-fastgpt-url", default=DEFAULT_BASE_FASTGPT_URL)
p.add_argument("--base-backend-url", default=DEFAULT_BASE_BACKEND_URL)
p.add_argument("--outer-backend-url", default=DEFAULT_OUTER_BACKEND_URL)
def build_arg_parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(description="上传、下载或补全 FastGPT/后端文件 URL。")
def _build_arg_parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(description="上传或下载文件。")
sub = p.add_subparsers(dest="command", required=True)
u = sub.add_parser("upload"); add_url_args(u); u.add_argument("--username", default=DEFAULT_BACKEND_ADMIN_USERNAME); u.add_argument("--password", default=DEFAULT_BACKEND_ADMIN_PASSWORD); u.add_argument("path")
d = sub.add_parser("download"); add_url_args(d); d.add_argument("url"); d.add_argument("path")
n = sub.add_parser("normalize-url"); add_url_args(n); n.add_argument("url")
u = sub.add_parser("upload"); _add_url_args(u); u.add_argument("--username", default=DEFAULT_BACKEND_ADMIN_USERNAME); u.add_argument("--password", default=DEFAULT_BACKEND_ADMIN_PASSWORD); u.add_argument("path")
d = sub.add_parser("download"); _add_url_args(d); d.add_argument("url"); d.add_argument("path", nargs="?")
return p
def main(argv: list[str] | None = None) -> int:
p = build_arg_parser(); a = p.parse_args(argv)
configure_urls(_strip(a.base_fastgpt_url), _strip(a.base_backend_url), _strip(a.outer_backend_url))
p = _build_arg_parser(); a = p.parse_args(argv)
_configure_urls(_strip(a.base_fastgpt_url), _strip(a.base_backend_url), _strip(a.outer_backend_url))
if a.command == "upload":
configure_login(a.username, a.password); print(upload_file(a.path)); return 0
_configure_login(a.username, a.password); print(upload_file(a.path)); return 0
if a.command == "download":
saved = download_file(a.url, a.path)
if saved is None:
return 1
print(saved); return 0
if a.command == "normalize-url":
print(url_replace_fastgpt(a.url)); return 0
p.error(f"unsupported command: {a.command}"); return 2
......
skills/ocr-skill/SKILL.md 0 → 100644
---
name: ocr-skill
description: OCR Skill。提供基于本地 Tesseract 的图片和 PDF 文本识别 CLI，以及可在 Python 中复用的 `TesseractOCRUtil` 类。
---
# OCR Skill
## 定位
`ocr-skill` 负责处理扫描件、图片和图片型 PDF 的文字识别。它适合在普通文本解析失败、PDF 文本乱码、合同是扫描版或截图版时使用。
该 Skill 只提供 OCR 能力，不负责合同审查、规则匹配、facts 提取或结果导出。识别出的文本可以继续交给 `doc-excel-skill`、`review-llm-skill` 或上层流程使用。
## 适用场景
- 识别合同截图或图片中的文字。
- 识别扫描版 PDF 每一页的文字。
- 在 PDF 直接解析结果为空或乱码时作为兜底方案。
- 在 Python 代码中直接调用 `TesseractOCRUtil` 做本地 OCR。
## 工具文件
- `scripts/ocr_tool.py`：OCR CLI 和 `TesseractOCRUtil` 类。
## 运行要求
- 本机需要安装 `tesseract` 可执行文件，并确保它在 `PATH` 中。
- 中文识别需要安装对应语言包，例如 `chi_sim`。
- PDF 转图片依赖 `PyMuPDF`，包名为 `PyMuPDF`，导入名为 `fitz`。
- OCR 质量受扫描清晰度、页眉页脚、表格线、印章和图片压缩影响。
## 主要命令
- `image`：识别单张图片，输出纯文本。
- `pdf`：把 PDF 每页转为图片后 OCR，输出每页识别结果 JSON。
## 输入输出
- 图片 OCR 输入图片路径，输出识别文本。
- PDF OCR 输入 PDF 路径，输出包含页码和文本的 JSON。
- 默认语言和 tesseract 可执行路径可通过命令参数覆盖，具体参数以 `--help` 为准。
## 使用示例
查看帮助：
```bash
python skills/ocr-skill/scripts/ocr_tool.py --help
```
识别图片：
```bash
python skills/ocr-skill/scripts/ocr_tool.py image demo/ocr.png
```
识别 PDF：
```bash
python skills/ocr-skill/scripts/ocr_tool.py pdf skills/ocr-skill/example/example.pdf
```
Python 中直接使用：
```python
from pathlib import Path
import sys
sys.path.append(str(Path("skills/ocr-skill/scripts").resolve()))
from ocr_tool import TesseractOCRUtil
util = TesseractOCRUtil(lang="chi_sim+eng", executable="tesseract")
text = util.ocr_image("/path/to/image.png")
print(text)
texts = util.ocr_result_pdf("/path/to/document.pdf")
print(texts)
```
## 在合同审查流程中的位置
该 Skill 通常作为文档解析阶段的兜底能力。当 `doc-excel-skill` 无法直接读取有效文本时，可以先用 OCR 得到页面文字，再进入分段、摘要、审查和导出流程。
skills/review-llm-skill/SKILL.md 0 → 100644
---
name: review-llm-skill
description: LLM 动作执行模块。按动作选择系统提示词，并把传入的 rule dict 作为用户提示词发送给模型。
---
# Review LLM Skill
## 定位
`review-llm-skill` 是一个可单独执行的 LLM 动作模块。
它只做两件事：
- 根据 `action` 选择对应的系统提示词。
- 将调用方传入的 `rule` dict 和待处理文本拼成 user prompt。
本模块不负责读取规则、不选择规则、不编排流程、不保存状态。
## 支持动作
- `summary` / `segment_summary` / `摘要` / `总结`
- `router` / `segment_rule_router` / `路由`
- `review` / `审查`
- `reflect` / `反思` / `复核`
- `merge` / `merger` / `segment_merger` / `合并`
## 工具文件
- `scripts/segment_llm_action.py`：主入口，负责动作调度和 LLM 调用。
- `scripts/prompts.py`：系统提示词。
- `scripts/llm_tool.py`：OpenAI 兼容 LLM 调用与 JSON 解析。
- `scripts/config.py`：LLM 配置。
## 输入
- `action`：要执行的动作。
- `--rule`：任意字段的 JSON dict，支持 `@file.json`。
- `--text`：直接输入待处理文本。
- `--input-file` + `--chunk-size` + `--chunk-index`：从文本文件中按字符数切片读取待处理文本。
- `--output`：输出目标；默认 `-` 表示直接打印，传入文件路径则追加到 JSON 数组文件。
## Python 接口
```python
from segment_llm_action import run_segment_llm_action
res = run_segment_llm_action(
action="review",
rule={
"title": "付款审查",
"rule": "检查付款期限是否明确",
"context": {"party_role": "甲方"},
},
text="甲方应于合同签订之日起30日内付款。",
)
print(res)
```
## CLI 示例
```bash
python skills/review-llm-skill/scripts/segment_llm_action.py review \
--rule '{"title":"付款审查","rule":"检查付款期限是否明确","context":{"party_role":"甲方"}}' \
--text '甲方应于合同签订之日起30日内付款。'
```
从文件读取指定分段：
```bash
python skills/review-llm-skill/scripts/segment_llm_action.py review \
--rule '{"title":"付款审查","rule":"检查付款期限是否明确"}' \
--input-file skills/review-llm-skill/example/downloaded_file.txt \
--chunk-size 2000 \
--chunk-index 0
```
追加输出到 JSON 文件：
```bash
python skills/review-llm-skill/scripts/segment_llm_action.py review \
--rule '{"title":"付款审查","rule":"检查付款期限是否明确"}' \
--input-file skills/review-llm-skill/example/downloaded_file.txt \
--chunk-size 2000 \
--chunk-index 0 \
--output outputs/review-results.json
```
只打印 messages，不调用模型：
```bash
python skills/review-llm-skill/scripts/segment_llm_action.py review \
--rule '{"title":"付款审查","rule":"检查付款期限是否明确"}' \
--text '甲方应于合同签订之日起30日内付款。' \
--print-messages
```
skills/review-llm-skill/example/downloaded_file.txt 0 → 100644
购销合同
供方：海南金盘智能科技股份有限公司 签订地点： 太原市
需方：山西长缘电力工程有限公司 签订时间： 2026年06月10日
一、货物（服务）名称、商标、型号、厂家、数量、金额 价格单位：（元）
货物名称
规格型号
生产厂家
单位
数量
单价（元）
总金额（元）
变压器
ZLSCLB-1000/10（6）
海南金盘
套
1
103100
103100
合计人民币金额： 大写 壹拾万零叁仟壹佰元整 小写：¥103100元
含：国标变压器本体、温控、IP20钢板外壳（标准色为RAL7035)、包装运输及13%增值税票等。
图号：DK1457.01.12GZ
技术参数：连接组别：Dyn11； 阻抗：6%；分接范围：±2×5% （变压器外壳与太重挖掘机全焊接抗震性相同）
二、交(提)货时间、地点：合同签订且方案或技术协议签订后 45 日内发到指定地点。
三、质量要求、技术标准：按国家及行业规范，产品交付之日起十八个月，或产品运行之日起十二个月，两者以先到时间为准。在保修(质保)期内如出现产品质量问题由卖方负责免费“三包”；操作、使用或保养不当等造成损坏的或不属产品质量问题的不在“三包”服务之列。
四、运输方式及到达站港和费用负担：由供方负担。
五、合理损耗及计算方法： 无损耗。
六、包装标准、包装物的供应与回收和费用负担：按国家及行业规范包装，包装物不回收。
七、验收标准、方法：按国家及行业规定。
八、异议期限及处理方法：需方收货后 3个月内或在货物安装使用后 6个月内发现货物存在质量问题，提出书面异议，双方协商解决。
九、随机备品、配件工具数量及供应方法：无备品配件，随货带装置说明书。
十、结算方式及期限：1、电汇或一线银行开具的6个月以内银行承兑汇票 2、合同签订后，发货前付清全款，供方开具税率为13%的增值税专用发票。
十一、违约责任：按中国法律。本合同双方签字盖章的扫描件具备与纸质版同等的法律效力。
十二、解决合同纠纷的方式：由双方友好协商；若协商不成则由卖方所在地法院管辖。
需 方
供 方
买受人（章）
山西长缘电力工程有限公司
出卖人（章）
海南金盘智能科技股份有限公司
地址：
山西省太原市小店区平阳路14号26幢20层2001、2002、2003号（太原首信商务秘书有限公司-1144号）集群登记
地址：
海南省海口市南海大道168-39号
法定代表人：
马林俊
法定代表人：
李辉
委托代理人：（签章）
委托代理人：（签章）
电话：
电话：
0898-66811301
开户银行：
中国农业银行太原平阳南路支行
开户银行：
交通银行海口南海支行
帐号：
04138201040004607
帐号：
461602303018010043627
税务登记号：
91140105MAENF9FL7F
税务登记号：
9146010062006446XN
邮政编码：
邮政编码：
skills/review-llm-skill/scripts/__init__.py deleted 100644 → 0
"""Standalone CLI scripts for review-llm-skill."""
skills/review-llm-skill/scripts/config.py
"""Compact prompt templates kept for compatibility."""
PROMPTS = {
"review": "基于当前分段和审查规则审查合同，仅输出JSON：{\"overall_conclusion\":\"\",\"findings\":[]}。\n分段：{segment_text}\n立场：{party_role}\n规则：{ruleset_text}",
"summary": "提取当前分段中与规则字段相关的客观事实，仅输出JSON：{\"facts\":{}}。\n分段：{segment_text}\n字段：{rule_fields}",
"router": "从候选规则中选择当前分段应执行的审查项，仅输出JSON：{\"selected_items\":[]}。\n分段：{segment_text}\n记忆：{context_memories_json}\n立场：{party_role}\n候选：{candidate_rules_json}",
"merger": "合并重复或相关的不合格findings，仅输出JSON：{\"findings\":[]}。\n输入：{payload}",
"reflect": "基于规则、已有findings和facts复核、去重、拆分、合并并定稿，仅输出JSON：{\"final_findings\":[]}。\n规则：{rule}\nfindings：{findings_json}\nfacts：{facts_json}\n立场：{party_role}",
"fact-merge": "合并summary_name下多个分段facts，不新增事实，仅输出JSON：{\"merge_facts\":{}}。\nsummary_names:{summary_names_json}\nfacts:{facts_json}",
"ruleset-route": "从候选ruleset_id中按问题选择一个，不得编造，仅输出JSON：{\"ruleset_id\":\"\",\"reason\":\"\"}。\n候选：{ruleset_ids_json}\n问题：{question}",
"party-role": "分析指定公司在合同中的商业角色，不仅按甲乙方判断，仅输出JSON：{\"party_role\":\"demand_side | supplier_side | unclear\",\"reason\":\"\"}。\n公司：{company_name}\n合同：{contract_text}",
"llm": "你是通用LLM助手。",
}
import os
OPENAI_MODEL = os.environ.get("OPENAI_MODEL", "Qwen3.5-122B-A10B-AWQ")
OPENAI_BASE_URL = os.environ.get("OPENAI_BASE_URL","http://192.168.252.71:9002/v1")
OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY","none")
USE_FASTGPT_SYSTEM_VARIABLE = False
DISABLE_LLM_THINKING = True
\ No newline at end of file
skills/review-llm-skill/scripts/llm_tool.py
import os
import re
import json
import urllib.request
from typing import Any, List, Dict
from openai import OpenAI
from tenacity import retry, stop_after_attempt, stop_after_delay, wait_fixed
try:
from .config import (
DISABLE_LLM_THINKING,
OPENAI_API_KEY,
OPENAI_BASE_URL,
OPENAI_MODEL,
USE_FASTGPT_SYSTEM_VARIABLE,
)
except ImportError:
from config import (
DISABLE_LLM_THINKING,
OPENAI_API_KEY,
OPENAI_BASE_URL,
OPENAI_MODEL,
USE_FASTGPT_SYSTEM_VARIABLE,
)
class LLMTool:
def __init__(self, system_prompt: str = ""):
self.system_prompt = system_prompt or ""
self.model = os.environ.get("OPENAI_MODEL", "gpt-4o-mini")
self.base_url = (os.environ.get("OPENAI_BASE_URL") or "https://api.openai.com/v1").rstrip("/")
self.api_key = os.environ.get("OPENAI_API_KEY")
self.model = OPENAI_MODEL
self.base_url = OPENAI_BASE_URL
self.api_key = OPENAI_API_KEY
self.client = OpenAI(base_url=self.base_url, api_key=self.api_key) if self.api_key else None
self.use_fastgpt_system_variable = USE_FASTGPT_SYSTEM_VARIABLE
self.disable_thinking = DISABLE_LLM_THINKING
def build_messages(self, user_content: str, system_content: str | None = None) -> List[Dict[str, str]]:
msgs = []
......@@ -19,22 +40,37 @@ class LLMTool:
msgs.append({"role": "user", "content": user_content})
return msgs
def _prepare_request(
self, messages: List[Dict[str, str]]
) -> tuple[List[Dict[str, str]], Dict[str, Any]]:
request_messages = list(messages)
extra_body: Dict[str, Any] = {}
if self.use_fastgpt_system_variable and request_messages and request_messages[0].get("role") == "system":
extra_body["variables"] = {"system": request_messages[0].get("content", "")}
request_messages = request_messages[1:]
if self.disable_thinking:
extra_body["thinking"] = {"type": "disabled"}
extra_body["chat_template_kwargs"] = {"enable_thinking": False}
return request_messages, extra_body
@retry(stop=stop_after_delay(600) | stop_after_attempt(3), wait=wait_fixed(1))
def run(self, messages: List[Dict[str, str]]) -> str:
if not self.api_key:
if not self.client:
raise RuntimeError("OPENAI_API_KEY is required")
body = json.dumps({"model": self.model, "messages": messages}, ensure_ascii=False).encode("utf-8")
request = urllib.request.Request(
f"{self.base_url}/chat/completions",
data=body,
headers={
"Content-Type": "application/json",
"Authorization": f"Bearer {self.api_key}",
},
method="POST",
)
with urllib.request.urlopen(request, timeout=120) as response:
payload = json.loads(response.read().decode("utf-8"))
return (((payload.get("choices") or [{}])[0].get("message") or {}).get("content")) or ""
request_messages, extra_body = self._prepare_request(messages)
kwargs: Dict[str, Any] = {
"model": self.model,
"messages": request_messages,
}
if extra_body:
kwargs["extra_body"] = extra_body
response = self.client.chat.completions.create(**kwargs)
return response.choices[0].message.content or ""
def chat_async(self, messages: List[Dict[str, str]]) -> str:
return self.run(messages)
......
skills/review-llm-skill/scripts/prompts.py 0 → 100644
"""System prompts for review-llm-skill."""
from __future__ import annotations
REVIEW_SYSTEM_PROMPT = """
你是一个专业的合同分段审查智能体（SegmentReview）。
你的任务是：基于给定审查项规则，对“待处理文本”进行审查，识别其中与规则相关且证据充分的条款，并判断其结果为“合格”或“不合格”，输出审查结论及必要的修改建议。
【输入说明】
- 审查项名称：本次要执行的审查项。
- 审查项规则：由 rules.xlsx 加载得到的该审查项规则 dict。
- 上下文信息：调用方提供的合同上下文、角色、事实、已有结论等 dict。
- 待处理文本：本次需要审查的文本分段。
【审查范围】
你只能审查待处理文本自身已经明确体现的内容。
你只能识别合格条款和不合格条款，不得对无关或证据不足内容生成 finding。
【审查原则】
- 严格基于给定审查项规则进行审查，不得脱离规则自行扩展审查标准。
- 可以读取上下文信息辅助理解主体、角色、术语和已有事实，但不得用上下文信息替代待处理文本中的证据。
- 优先识别“确定成立”的合格或不合格结论，不输出模糊怀疑类表述。
- 必须逐句扫描待处理文本，穷举所有证据充分的问题或合格表述。
【单一证据约束】
每一个 finding 必须只对应一个独立判断点和一个最小证据句；若多个句子分别支持不同问题，必须拆分为多个 findings；严禁在 original_text 中拼接多个不连续句子。
【输出约束】
严格输出 JSON 数组；不得输出 JSON 之外的解释性文字。若未发现证据充分的合格或不合格条款，返回 []。
【输出格式】
[
{
"rule_name": "审查项名称",
"result": "合格 或 不合格",
"issue": "基于规则和原文说明为什么合格或不合格",
"original_text": "待处理文本中的最小证据原文",
"suggestion": "合格时填写“无需修改”；不合格时填写具体、可执行的修改建议"
}
]
"""
REFLECT_SYSTEM_PROMPT = """
你是一个合同审查反思智能体（ReviewReflection）。
你的任务不是从零重新审查合同，也不是简单删减 findings，而是基于“审查项规则、待处理文本、上下文信息中的已有 findings/facts/角色/全文信息”，对已有 findings 进行规则内复核、去重、校正、拆分、合并与定稿，输出最终 findings 数组。
【输入说明】
- 审查项名称：本次要反思复核的审查项。
- 审查项规则：由 rules.xlsx 加载得到的该审查项规则 dict。
- 上下文信息：调用方提供的已有 findings、facts、合同全文、合同角色等 dict。
- 待处理文本：本次复核对应的文本，可以是分段文本或相关全文片段。
【允许执行的操作】
删除重复、证据不足、引用不当或超出当前审查项规则的 findings；修订 issue、result、original_text 或 suggestion 不准确的 findings；合并多个指向同一问题的 findings；拆分包含多个独立问题的 finding。
【禁止事项】
不得脱离当前审查项规则新增全新的审查维度；不得凭空创造合同中不存在的事实；不得输出无法由合同原文直接支持的结论；不得输出模糊、空泛、不可执行的 suggestion。
【核心判定原则】
final result 必须以“审查项规则 + 待处理文本 + 上下文信息”为准；每条 final finding 必须能被合同原文直接支持；original_text 必须是最小充分证据片段；result 只能为“合格”或“不合格”。
【输出约束】
严格输出 JSON 数组；不得输出任何解释性文字；若反思后无成立 findings，返回 []。
【输出格式】
[
{
"rule_name": "审查项名称",
"result": "合格 或 不合格",
"issue": "复核后的准确风险或合格说明",
"original_text": "合同原文中的最小证据片段",
"suggestion": "可直接替换原文、新增条款措辞，或明确的修改方向"
}
]
"""
SUMMARY_SYSTEM_PROMPT = """
你是合同事实提取智能体（SegmentSummary）。
你的任务是：**基于给定的审查规则，从当前合同分段中提取“与该规则直接相关的客观事实”，并结构化输出。**
【核心原则】
你必须严格围绕“规则所需信息”进行提取。
---
【事实定义】
事实必须满足：
1. 可以在当前分段原文中直接找到对应表述；
2. 不得对原文进行抽象、概括或推断；
3. 不得补充未出现的主体、条件或数值；
4. 允许对原文做最小结构化拆分（例如金额、比例、期限）。
---
【规则驱动提取要求（关键）】
- 仅提取“该审查规则执行所需要的信息字段”
- 不得提取与该规则无关的信息（即使这些信息在文本中存在）
- 若规则未涉及某类信息，则不得输出对应字段
- 若规则涉及某字段但文本未出现，需显式标记为 "未明确"
---
【输出结构】
- 输出字段：facts
- facts 是一个对象
- 键必须来自【规则字段定义（rule_fields）】
- 不得使用预设通用维度（如“支付/违约责任”等）
---
【字段填充规则】
- 每个字段值必须是对象或对象列表
- 不得输出字符串作为字段值
- 字段内容必须为原文的最小结构化表达
- 不得改写原文含义
---
【缺失信息处理（非常重要）】
- 若规则要求的字段在当前分段未出现：
→ 必须输出该字段，并标记为：
"未明确"
（用于后续审查判断）
---
【约束】
- 严禁编造信息
- 严禁推断未出现的内容
- 不得输出风险判断或解释
- 严格输出 JSON
【输出格式示例】
```json
{
"facts": {
"支付审查": {"方式": "银行转账", "时间": "验收后30日内"},
"违约责任审查": {"违约金比例": "合同总金额的5%"}
}
}
```
"""
ROUTER_SYSTEM_PROMPT = """
你是合同分段规则路由智能体（SegmentRuleRouter）。
你的任务是：基于“当前分段文本”，从候选审查规则中选出“应执行审查”的规则项。
【路由目标】
- 仅做规则适配判断，不输出风险结论、不输出审查建议。
- 高召回优先：只要当前分段与规则存在明确相关性，就应路由命中。
- 若候选规则明显无关，则不要命中。
【判断依据】
- 以当前分段文本为主。
- 可参考上下文记忆辅助理解术语，但不得脱离当前分段文本做臆断。
【输出约束】
- 严格输出 JSON。
- 每个命中规则需给出简短 reason，说明该分段为何与规则相关。
- 若确实没有任何相关规则，返回 {"selected_items": []}。
【输出格式示例】
```json
{
"selected_items": [
{
"title": "规则标题",
"reason": "命中原因（简短）"
}
]
}
```
"""
MERGE_SYSTEM_PROMPT = """
你将收到同一组 findings 的 issue 与 suggestion 列表，请做信息融合而非机械拼接。
要求：
1. 输入中已经包含同组条款原文`original_text`，请仅将其作为分析依据。
2. `issue`：提炼并合并组内风险点，去重、保留关键信息，语言精炼。
3. `suggestion`：合并为一条可执行建议，必须基于输入原文的具体表述来给出，避免空泛、泛化或与原文脱节，必要时按“先补充条款、再明确标准”这类逻辑组织。
4. 禁止输出与输入无关的信息。
【输出格式示例】
```json
{
"issue": "提炼合并后的风险点",
"suggestion": "提炼合并后的建议"
}
```
"""
skills/review-llm-skill/scripts/review_llm_skill.py
#!/usr/bin/env python3
"""Standalone review LLM CLI."""
"""Backward-compatible entry point for segment_llm_action.py."""
from __future__ import annotations
import argparse, json, os, re, urllib.request
from pathlib import Path
from typing import Any
PROMPTS = {
"review": "基于当前分段和审查规则审查合同，仅输出JSON：{\"overall_conclusion\":\"\",\"findings\":[]}。\n分段：{segment_text}\n立场：{party_role}\n规则：{ruleset_text}",
"summary": "仅提取当前分段中与规则字段相关的客观事实，仅输出JSON：{\"facts\":{}}。\n分段：{segment_text}\n字段：{rule_fields}",
"router": "从候选规则中选择当前分段应执行的审查项，仅输出JSON：{\"selected_items\":[]}。\n分段：{segment_text}\n记忆：{context_memories_json}\n立场：{party_role}\n候选：{candidate_rules_json}",
"merger": "合并同一分段内重复或相关的不合格findings，仅输出JSON：{\"findings\":[]}。\n输入：{payload}",
"reflect": "基于规则、已有findings和facts复核、去重、拆分、合并并定稿，仅输出JSON：{\"final_findings\":[]}。\n规则：{rule}\nfindings：{findings_json}\nfacts：{facts_json}\n立场：{party_role}",
"fact-merge": "合并summary_name下多个分段facts，不新增事实，仅输出JSON：{\"merge_facts\":{}}。\nsummary_names:{summary_names_json}\nfacts:{facts_json}",
"ruleset-route": "从候选ruleset_id中按问题选择一个，不得编造，仅输出JSON：{\"ruleset_id\":\"\",\"reason\":\"\"}。\n候选：{ruleset_ids_json}\n问题：{question}",
"party-role": "分析指定公司在合同中的商业角色，不仅按甲乙方判断，仅输出JSON：{\"party_role\":\"demand_side | supplier_side | unclear\",\"reason\":\"\"}。\n公司：{company_name}\n合同：{contract_text}",
"llm": "你是通用LLM助手。",
}
class LLMTool:
def __init__(self, system_prompt: str = ""):
self.system_prompt = system_prompt
self.model = os.environ.get("OPENAI_MODEL", "gpt-4o-mini")
self.base_url = (os.environ.get("OPENAI_BASE_URL") or "https://api.openai.com/v1").rstrip("/")
self.api_key = os.environ.get("OPENAI_API_KEY")
def build_messages(self, user: str, system: str | None = None) -> list[dict[str, str]]:
return ([{"role": "system", "content": system}] if system else []) + [{"role": "user", "content": user}]
def run(self, messages: list[dict[str, str]]) -> str:
if not self.api_key:
raise RuntimeError("OPENAI_API_KEY is required")
body = json.dumps({"model": self.model, "messages": messages}, ensure_ascii=False).encode()
req = urllib.request.Request(f"{self.base_url}/chat/completions", data=body, headers={"Content-Type": "application/json", "Authorization": f"Bearer {self.api_key}"}, method="POST")
with urllib.request.urlopen(req, timeout=120) as resp:
data = json.loads(resp.read().decode())
return (((data.get("choices") or [{}])[0].get("message") or {}).get("content")) or ""
def parse_first_json(self, text: str) -> Any:
if not text:
return None
try:
return json.loads(text)
except Exception:
m = re.search(r"(\{.*\}|\[.*\])", text, re.S)
if not m:
return None
try:
return json.loads(m.group(1))
except Exception:
return None
def jdump(v: Any) -> str:
return json.dumps(v, ensure_ascii=False, indent=2)
def rules_text(rules: list[dict[str, Any]]) -> str:
return "\n".join(f"标题:{r.get('title','')}\n规则:{r.get('rule','')}\n等级:{r.get('level','')}\n建议:{r.get('suggestion_template','')}\n案例:{r.get('case','')}" for r in rules or [])
def default_rulesets() -> list[str]:
return ["合同信息提取（合同组）", "合同信息提取（技术部）", "合同信息提取（采购部）", "技术协议提取（合同组）", "技术协议提取（技术部）"]
def empty_fact(v: Any) -> bool:
return v is None or (isinstance(v, str) and (not v.strip() or v.strip() == "未明确")) or (isinstance(v, (dict, list)) and not v)
def dedupe(values: list[Any]) -> list[Any]:
out, seen = [], set()
for v in values:
key = jdump(v) if isinstance(v, (dict, list)) else str(v)
if not empty_fact(v) and key not in seen:
seen.add(key); out.append(v)
return out
def merge_facts_rule(facts: list[dict[str, Any]], names: list[str]) -> dict[str, Any]:
merged: dict[str, Any] = {}
for name in dict.fromkeys(str(n).strip() for n in names or [] if str(n).strip()):
vals = dedupe([item.get(name) for item in facts or [] if isinstance(item, dict) and name in item])
if vals:
merged[name] = vals[0] if len(vals) == 1 else vals
merged["_meta"] = {"summary_names": names, "source_fact_count": len(facts or [])}
return merged
def run_review_llm(tool_name: str = "review", segment_id: int = 0, user_prompt: str | None = None, **kw) -> dict[str, Any]:
name = (tool_name or "review").lower()
if name == "fact-merge" and str(kw.get("merge_mode") or "llm").lower() != "llm":
return {"merge_facts": merge_facts_rule(kw.get("facts") or [], kw.get("summary_names") or [])}
if name == "llm":
prompt, user = user_prompt or PROMPTS["llm"], kw.get("user_content") or kw.get("segment_text") or ""
elif name == "review":
prompt, user = PROMPTS[name].format(segment_text=kw.get("segment_text", ""), party_role=kw.get("party_role", ""), ruleset_text=rules_text(kw.get("rules") or [])), ""
elif name == "summary":
prompt, user = PROMPTS[name].format(segment_text=kw.get("segment_text", ""), rule_fields=jdump([r.get("summary") for r in kw.get("rules") or [] if r.get("summary")])), ""
elif name == "router":
prompt, user = PROMPTS[name].format(segment_text=kw.get("segment_text", ""), context_memories_json=jdump(kw.get("context_facts") or []), party_role=kw.get("party_role", ""), candidate_rules_json=jdump([{r.get("title", ""): r.get("rule", "")} for r in kw.get("rules") or []])), ""
elif name == "merger":
prompt, user = PROMPTS[name].format(payload=jdump(kw.get("payload") or kw.get("findings") or [])), ""
elif name == "reflect":
prompt, user = PROMPTS[name].format(rule=jdump(kw.get("rule") or {}), findings_json=jdump(kw.get("findings") or []), facts_json=jdump(kw.get("facts") or kw.get("context_facts") or []), party_role=kw.get("party_role", "")), ""
elif name == "fact-merge":
prompt, user = PROMPTS[name].format(summary_names_json=jdump(kw.get("summary_names") or []), facts_json=jdump(kw.get("facts") or [])), ""
elif name == "ruleset-route":
prompt, user = PROMPTS[name].format(question=kw.get("question") or user_prompt or "", ruleset_ids_json=jdump(kw.get("ruleset_ids") or default_rulesets())), ""
elif name == "party-role":
prompt, user = PROMPTS[name].format(company_name=kw.get("company_name") or "", contract_text=kw.get("contract_text") or kw.get("segment_text") or ""), ""
else:
return {"error": f"unknown tool: {tool_name}"}
llm = LLMTool(prompt); raw = llm.run(llm.build_messages(user, prompt)); return llm.parse_first_json(raw) or {"raw": raw}
def load_arg(v: str | None, default: Any) -> Any:
if v is None:
return default
return json.loads(Path(v[1:]).read_text(encoding="utf-8") if v.startswith("@") else v)
def parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(description="Standalone review LLM CLI")
p.add_argument("tool_name", nargs="?", default="review", choices=["review", "summary", "router", "merger", "reflect", "fact-merge", "ruleset-route", "party-role", "llm"])
for name, default in [("segment-id", 0), ("segment-text", ""), ("segment-text-file", None), ("party-role", ""), ("rules", "[]"), ("context-facts", "{}"), ("payload", None), ("findings", "[]"), ("facts", "[]"), ("rule", "{}"), ("summary-names", "[]"), ("question", ""), ("ruleset-ids", None), ("company-name", ""), ("contract-text", ""), ("contract-text-file", None), ("user-prompt", None), ("user-content", None)]:
flag = f"--{name}"; kwargs = {"default": default}
if name == "segment-id": kwargs["type"] = int
p.add_argument(flag, **kwargs)
p.add_argument("--merge-mode", choices=["llm", "rule"], default="llm"); p.add_argument("--output-raw", action="store_true")
return p
def main(argv: list[str] | None = None) -> int:
a = parser().parse_args(argv)
seg = Path(a.segment_text_file).read_text(encoding="utf-8") if a.segment_text_file else a.segment_text
contract = Path(a.contract_text_file).read_text(encoding="utf-8") if a.contract_text_file else a.contract_text
kw = {"segment_text": seg, "party_role": a.party_role, "rules": load_arg(a.rules, []), "context_facts": load_arg(a.context_facts, {}), "payload": load_arg(a.payload, []) if a.payload else None, "findings": load_arg(a.findings, []), "facts": load_arg(a.facts, []), "rule": load_arg(a.rule, {}), "summary_names": load_arg(a.summary_names, []), "merge_mode": a.merge_mode, "question": a.question, "ruleset_ids": load_arg(a.ruleset_ids, default_rulesets()) if a.ruleset_ids else default_rulesets(), "company_name": a.company_name, "contract_text": contract, "user_content": a.user_content}
res = run_review_llm(a.tool_name, a.segment_id, a.user_prompt, **kw)
print(res if a.output_raw else jdump(res))
return 0
from segment_llm_action import * # noqa: F403
from segment_llm_action import main
if __name__ == "__main__":
......
skills/review-llm-skill/scripts/segment_llm_action.py 0 → 100644
#!/usr/bin/env python3
"""Run an LLM action with a rule payload."""
from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
from typing import Any
SCRIPT_DIR = Path(__file__).resolve().parent
if str(SCRIPT_DIR) not in sys.path:
sys.path.insert(0, str(SCRIPT_DIR))
try:
from .prompts import (
MERGE_SYSTEM_PROMPT,
REFLECT_SYSTEM_PROMPT,
REVIEW_SYSTEM_PROMPT,
ROUTER_SYSTEM_PROMPT,
SUMMARY_ROUTER_SYSTEM_PROMPT,
SUMMARY_SYSTEM_PROMPT,
)
except ImportError:
from prompts import (
MERGE_SYSTEM_PROMPT,
REFLECT_SYSTEM_PROMPT,
REVIEW_SYSTEM_PROMPT,
ROUTER_SYSTEM_PROMPT,
SUMMARY_ROUTER_SYSTEM_PROMPT,
SUMMARY_SYSTEM_PROMPT,
)
Action = str
def jdump(value: Any) -> str:
return json.dumps(value, ensure_ascii=False, indent=2)
def load_json_arg(value: str | None, default: Any) -> Any:
if value is None:
return default
if value.startswith("@"):
return json.loads(Path(value[1:]).read_text(encoding="utf-8"))
return json.loads(value)
def pick_text_chunk(text: str, chunk_size: int | None, chunk_index: int) -> str:
if chunk_size is None:
return text
if chunk_size <= 0:
raise ValueError("--chunk-size must be > 0")
if chunk_index < 0:
raise ValueError("--chunk-index must be >= 0")
start = chunk_index * chunk_size
return text[start : start + chunk_size]
def load_review_text(text: str, input_file: str | None, chunk_size: int | None, chunk_index: int) -> str:
if input_file:
file_text = Path(input_file).read_text(encoding="utf-8")
return pick_text_chunk(file_text, chunk_size, chunk_index)
return text
def system_prompt_for(action: Action) -> str:
return {
"review": REVIEW_SYSTEM_PROMPT,
"reflect": REFLECT_SYSTEM_PROMPT,
"summary": SUMMARY_SYSTEM_PROMPT,
"router": ROUTER_SYSTEM_PROMPT,
"merge": MERGE_SYSTEM_PROMPT,
}[action]
def ensure_rule_dict(rule: Any) -> dict[str, Any]:
if not isinstance(rule, dict):
raise ValueError("--rule must be a JSON object")
return rule
def build_user_prompt(rule: dict[str, Any], text: str) -> str:
return jdump({"rule": ensure_rule_dict(rule), "text": text})
def build_messages(action: Action, rule: dict[str, Any], text: str = "") -> list[dict[str, str]]:
normalized_action = normalize_action(action)
return [
{"role": "system", "content": system_prompt_for(normalized_action)},
{"role": "user", "content": build_user_prompt(rule, text)},
]
def append_json_output(output_file: str, value: Any) -> None:
path = Path(output_file)
if path.exists() and path.stat().st_size > 0:
data = json.loads(path.read_text(encoding="utf-8"))
if not isinstance(data, list):
raise ValueError(f"--output file must contain a JSON array: {output_file}")
else:
data = []
if isinstance(value, list):
data.extend(value)
else:
data.append(value)
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(jdump(data) + "\n", encoding="utf-8")
def write_output(output: str, value: Any) -> None:
if output == "-":
print(jdump(value))
return
append_json_output(output, value)
def create_llm_tool(system_prompt: str):
try:
from .llm_tool import LLMTool
except ImportError:
from llm_tool import LLMTool
return LLMTool(system_prompt)
def normalize_action(action: str) -> Action:
value = (action or "review").strip().lower()
aliases = {
"review": "review",
"审查": "review",
"reflect": "reflect",
"reflection": "reflect",
"反思": "reflect",
"复核": "reflect",
"summary": "summary",
"segment_summary": "summary",
"summarize": "summary",
"摘要": "summary",
"总结": "summary",
"router": "router",
"route": "router",
"segment_rule_router": "router",
"路由": "router",
"merge": "merge",
"merger": "merge",
"segment_merger": "merge",
"合并": "merge",
}
if value not in aliases:
raise ValueError(f"unknown action: {action}")
return aliases[value]
def run_segment_llm_action(action: Action, rule: dict[str, Any], text: str = "") -> Any:
messages = build_messages(action, rule, text)
llm = create_llm_tool(messages[0]["content"])
raw = llm.run(messages)
return llm.parse_first_json(raw) or {"raw": raw}
def parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(description="Run an LLM action with a rule payload")
p.add_argument("action", choices=["review", "reflect", "summary", "segment_summary", "router", "segment_rule_router", "merge", "merger", "segment_merger", "审查", "反思", "复核", "摘要", "总结", "路由", "摘要路由", "摘要项路由", "合并"])
p.add_argument("--rule", required=True, help="任意字段的 JSON dict；支持 @file.json")
p.add_argument("--text", "--segment-text", dest="text", default="", help="直接输入待审查文本")
p.add_argument("--input-file", help="从文本文件读取待审查文本")
p.add_argument("--chunk-size", type=int, help="从文件读取时的分段大小")
p.add_argument("--chunk-index", type=int, default=0, help="从文件读取时的分段序号，从 0 开始")
p.add_argument("--output", default="-", help="输出目标；'-' 直接打印，其他路径则追加到 JSON 数组文件")
p.add_argument("--print-messages", action="store_true", help="只打印构造后的 messages，不调用模型")
return p
def main(argv: list[str] | None = None) -> int:
args = parser().parse_args(argv)
rule = ensure_rule_dict(load_json_arg(args.rule, {}))
text = load_review_text(args.text, args.input_file, args.chunk_size, args.chunk_index)
messages = build_messages(args.action, rule, text)
if args.print_messages:
print(jdump(messages))
return 0
llm = create_llm_tool(messages[0]["content"])
raw = llm.run(messages)
write_output(args.output, llm.parse_first_json(raw) or {"raw": raw})
return 0
if __name__ == "__main__":
raise SystemExit(main())
skills/review-memory-skill/scripts/review_memory_cli.py deleted 100644 → 0
#!/usr/bin/env python3
"""Standalone JSON-backed review memory CLI."""
from __future__ import annotations
import argparse, json, logging
from dataclasses import asdict, dataclass
from pathlib import Path
from threading import RLock
from typing import Any
from uuid import uuid4
logging.basicConfig(level=logging.WARNING)
logger = logging.getLogger("review_memory_cli")
@dataclass
class Finding:
rule_title: str
segment_id: int
original_text: str
issue: str
risk_level: str
suggestion: str
id: str = ""
result: str = ""
@classmethod
def from_dict(cls, data: dict) -> "Finding":
d = data or {}
return cls(str(d.get("rule_title", "")), int(d.get("segment_id", 0) or 0), str(d.get("original_text", "")), str(d.get("issue", "")), str(d.get("risk_level", "")), str(d.get("suggestion", "")), str(d.get("id", "")), str(d.get("result", "")))
def to_dict(self) -> dict[str, Any]:
return asdict(self)
class MemoryStore:
def __init__(self, storage_name: str = "default.json") -> None:
self._storage_path = Path(__file__).resolve().parent.parent / "tmp" / storage_name
self._storage_path.parent.mkdir(parents=True, exist_ok=True)
self._lock = RLock()
self.facts: list[dict[str, Any]] = []
self.merge_facts: list[dict[str, Any]] = []
self.findings: dict[str, list[Finding]] = {}
self._load()
def _key(self, key: str | None) -> str:
return (key or "").strip().lower() or "review"
def add_fact(self, value: dict[str, Any]) -> list[dict[str, Any]]:
with self._lock:
self.facts.append(value); self._persist(); return self.facts
def add_merge_fact(self, value: dict[str, Any]) -> list[dict[str, Any]]:
with self._lock:
self.merge_facts.append(value); self._persist(); return self.merge_facts
def get_facts(self) -> list[dict[str, Any]]:
with self._lock:
return list(self.facts)
def add_finding(self, key: str, finding: Finding) -> Finding:
with self._lock:
if not finding.id:
finding.id = uuid4().hex
self.findings.setdefault(self._key(key), []).append(finding)
self._persist(); return finding
def list_findings(self, key: str | None = None) -> dict[str, list[dict[str, Any]]]:
with self._lock:
keys = [self._key(key)] if key else list(self.findings)
return {k: [f.to_dict() for f in self.findings.get(k, [])] for k in keys}
def get_findings_by_segment(self, key: str, segment_id: int) -> list[dict[str, Any]]:
return [f.to_dict() for f in self.findings.get(self._key(key), []) if f.segment_id == segment_id]
def search_findings(self, key: str, rule_title: str = "") -> list[dict[str, Any]]:
title = (rule_title or "").strip().lower()
return [f.to_dict() for f in self.findings.get(self._key(key), []) if not title or f.rule_title.lower() == title]
def delete_findings_by_segment(self, key: str, segment_id: int) -> int:
with self._lock:
k, current = self._key(key), list(self.findings.get(self._key(key), []))
self.findings[k] = [f for f in current if f.segment_id != segment_id]
removed = len(current) - len(self.findings[k])
if removed:
self._persist()
return removed
def search_facts(self, keywords: list[str]) -> list[Any]:
keys = [str(k).strip().lower() for k in keywords if str(k).strip()]
out = []
for item in self.facts:
for name, value in item.items():
low = str(name).lower()
if any(k in low or low in k for k in keys):
out.append({name: value})
return out
def clear(self) -> None:
with self._lock:
self.facts.clear(); self.merge_facts.clear(); self.findings.clear(); self._persist()
def _payload(self) -> dict[str, Any]:
return {"facts": self.facts, "merge_facts": self.merge_facts, "findings": {k: [f.to_dict() for f in v] for k, v in self.findings.items()}}
def _persist(self) -> None:
self._storage_path.write_text(json.dumps(self._payload(), ensure_ascii=False, indent=2), encoding="utf-8")
def _load(self) -> None:
if not self._storage_path.exists():
return
try:
data = json.loads(self._storage_path.read_text(encoding="utf-8") or "{}")
self.facts = data.get("facts") or []; self.merge_facts = data.get("merge_facts") or []
self.findings = {self._key(k): [Finding.from_dict(i) for i in items or []] for k, items in (data.get("findings") or {}).items()}
except Exception as exc:
logger.error("Failed to load memory store: %s", exc)
def export_to_json(self, path: str | None = None) -> str:
out = path or str(self._storage_path).replace(".json", "_export.json")
Path(out).write_text(json.dumps(self._payload(), ensure_ascii=False, indent=2), encoding="utf-8")
return out
def out(obj: Any) -> None:
print(json.dumps(obj, ensure_ascii=False, indent=2))
def load_json_arg(value: str) -> Any:
return json.loads(Path(value[1:]).read_text(encoding="utf-8") if value.startswith("@") else value)
def parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(prog="review-memory-cli"); p.add_argument("--storage", default="default.json")
sub = p.add_subparsers(dest="cmd")
sub.add_parser("list-facts")
a = sub.add_parser("add-fact"); a.add_argument("data")
a = sub.add_parser("add-merge-fact"); a.add_argument("data")
a = sub.add_parser("search-facts"); a.add_argument("keywords", nargs="+")
a = sub.add_parser("add-finding"); a.add_argument("--key", default="review"); a.add_argument("--rule", required=True); a.add_argument("--segment", type=int, default=0); a.add_argument("--original", default=""); a.add_argument("--issue", default=""); a.add_argument("--risk", default=""); a.add_argument("--suggest", default=""); a.add_argument("--result", default="")
sub.add_parser("list-findings")
a = sub.add_parser("list-findings-key"); a.add_argument("key")
a = sub.add_parser("findings-by-seg"); a.add_argument("key"); a.add_argument("segment", type=int)
a = sub.add_parser("search-findings"); a.add_argument("key"); a.add_argument("--rule-title", default="")
a = sub.add_parser("delete-findings-seg"); a.add_argument("key"); a.add_argument("segment", type=int)
sub.add_parser("clear")
a = sub.add_parser("export"); a.add_argument("--out")
return p
def main(argv: list[str] | None = None) -> int:
a = parser().parse_args(argv); store = MemoryStore(a.storage)
if a.cmd == "list-facts": out(store.get_facts()); return 0
if a.cmd == "add-fact": store.add_fact(load_json_arg(a.data)); print("OK"); return 0
if a.cmd == "add-merge-fact": store.add_merge_fact(load_json_arg(a.data)); print("OK"); return 0
if a.cmd == "search-facts": out(store.search_facts(a.keywords)); return 0
if a.cmd == "add-finding":
out(store.add_finding(a.key, Finding(a.rule, a.segment, a.original, a.issue, a.risk, a.suggest, result=a.result)).to_dict()); return 0
if a.cmd == "list-findings": out(store.list_findings()); return 0
if a.cmd == "list-findings-key": out(store.list_findings(a.key)); return 0
if a.cmd == "findings-by-seg": out(store.get_findings_by_segment(a.key, a.segment)); return 0
if a.cmd == "search-findings": out(store.search_findings(a.key, a.rule_title)); return 0
if a.cmd == "delete-findings-seg": print(store.delete_findings_by_segment(a.key, a.segment)); return 0
if a.cmd == "clear": store.clear(); print("cleared"); return 0
if a.cmd == "export": print(store.export_to_json(a.out)); return 0
parser().print_help(); return 1
if __name__ == "__main__":
raise SystemExit(main())
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or sign in to comment