mirror of
https://github.com/primedigitaltech/market-assistant.git
synced 2026-07-22 08:01:34 +08:00
feat(报告): 各章大模型衔接分析与 report_config 开关
Made-with: Cursor
This commit is contained in:
hub-gif
parent
b7d932b920
commit
a8210ec933
@ -33,3 +33,6 @@ CSRF_TRUSTED_ORIGINS=http://localhost:5173,http://127.0.0.1:5173
|
||||
# MA_SKIP_LLM_KEYWORD_SUGGEST=1
|
||||
# MA_ENABLE_LLM_COMMENT_SENTIMENT=1
|
||||
# MA_SKIP_LLM_COMMENT_SENTIMENT=1
|
||||
# 各章标题后插入大模型「衔接分析」（任务 report_config.llm_section_bridges 或本项设为 1）
|
||||
# MA_ENABLE_LLM_SECTION_BRIDGES=1
|
||||
# MA_SKIP_LLM_SECTION_BRIDGES=1
|
||||
|
||||
@ -1544,6 +1544,7 @@ def build_competitor_markdown(
|
||||
"- **评价主题词**：对评价正文做**预设词表子串计数**，非分词主题模型，适合扫方向，**需抽样人工验证**。",
|
||||
"- **用途/场景**：对每条评价独立判断是否命中预设场景词；一条可计入多个场景，统计的是「提及该场景的评价条数」而非用户数。",
|
||||
"- **用户画像（第八章）**：正负面粗判含**口语短语**级摘录；关注词与场景**仅按细类**以条形图展示（场景图为**占该细类有效文本比例 %**）；见 §8.3～8.4。",
|
||||
"- **各章衔接（可选）**：若任务配置 ``llm_section_bridges``（或部署侧环境变量启用），则在「## 一」至「## 九」各章二级标题后插入大模型撰写的**衔接分析**段落，便于阅读过渡；**定量结论仍以正文表格与摘要 JSON 为准**。",
|
||||
"- **检索结果规模**：来自京东 PC 搜索返回的「结果条数」类指标，表示平台侧申报的匹配数量级，**不等于**动销、库存或独立 SKU 数。",
|
||||
"",
|
||||
"### 1.4 主要局限",
|
||||
|
||||
@ -69,6 +69,25 @@ def merge_llm_report_with_rules_charts(llm_md: str, rules_md: str) -> str:
|
||||
return merge_llm_supplement_with_rules_report(llm_md, rules_md)
|
||||
|
||||
|
||||
def inject_section_bridges_into_markdown(md: str, bridges: dict[str, str]) -> str:
|
||||
"""
|
||||
在「## 一、」…「## 九、」各章标题行之后插入 ``#### 衔接分析（大模型）`` 段落。
|
||||
自第九章向前替换，避免多次插入导致偏移错位。
|
||||
"""
|
||||
out = md
|
||||
for key in "九八七六五四三二一":
|
||||
content = (bridges.get(key) or "").strip()
|
||||
if not content:
|
||||
continue
|
||||
pat = re.compile(rf"^(## {key}、[^\n]*)\n", re.MULTILINE)
|
||||
|
||||
def _repl(m: re.Match[str], _c: str = content) -> str:
|
||||
return m.group(0) + "\n#### 衔接分析（大模型）\n\n" + _c + "\n\n"
|
||||
|
||||
out = pat.sub(_repl, out, count=1)
|
||||
return out
|
||||
|
||||
|
||||
def _flat_comment_texts(comment_rows: list[dict[str, str]]) -> list[str]:
|
||||
"""全部非空评价正文（与报告统计同源）。"""
|
||||
out: list[str] = []
|
||||
@ -148,6 +167,7 @@ def get_default_report_config() -> dict[str, Any]:
|
||||
jcr, _ = _jd_crawler_modules()
|
||||
return {
|
||||
"llm_comment_sentiment": False,
|
||||
"llm_section_bridges": False,
|
||||
"comment_focus_words": list(jcr.COMMENT_FOCUS_WORDS),
|
||||
"comment_scenario_groups": [
|
||||
{"label": lbl, "triggers": list(trs)}
|
||||
@ -326,6 +346,49 @@ def write_competitor_analysis_for_run_dir(
|
||||
report_config=eff_rc,
|
||||
llm_sentiment_section_md=llm_sentiment_md or None,
|
||||
)
|
||||
|
||||
bridge_record: dict[str, Any] = {
|
||||
"schema_version": 1,
|
||||
"attempted": False,
|
||||
}
|
||||
skip_bridge = os.environ.get(
|
||||
"MA_SKIP_LLM_SECTION_BRIDGES", ""
|
||||
).strip().lower() in ("1", "true", "yes")
|
||||
env_bridge = os.environ.get(
|
||||
"MA_ENABLE_LLM_SECTION_BRIDGES", ""
|
||||
).strip().lower() in ("1", "true", "yes")
|
||||
want_bridge = bool(eff_rc.get("llm_section_bridges")) or env_bridge
|
||||
if want_bridge and not skip_bridge:
|
||||
from .llm_generate import (
|
||||
generate_section_bridges_llm,
|
||||
split_competitor_report_for_bridges,
|
||||
)
|
||||
|
||||
parts = split_competitor_report_for_bridges(md)
|
||||
if parts:
|
||||
bridge_record["attempted"] = True
|
||||
try:
|
||||
bridges = generate_section_bridges_llm(
|
||||
keyword=kw, brief=brief_final, sections=parts
|
||||
)
|
||||
bridge_record["keys_received"] = sorted(bridges.keys())
|
||||
md = inject_section_bridges_into_markdown(md, bridges)
|
||||
bridge_record["ok"] = True
|
||||
except Exception as e:
|
||||
bridge_record["ok"] = False
|
||||
bridge_record["error"] = str(e)
|
||||
else:
|
||||
bridge_record["skipped"] = "no_h2_sections_matched"
|
||||
elif skip_bridge:
|
||||
bridge_record["skipped"] = "MA_SKIP_LLM_SECTION_BRIDGES"
|
||||
elif not want_bridge:
|
||||
bridge_record["skipped"] = "not_enabled"
|
||||
|
||||
(run_dir / "section_bridge_llm.json").write_text(
|
||||
json.dumps(bridge_record, ensure_ascii=False, indent=2),
|
||||
encoding="utf-8",
|
||||
)
|
||||
|
||||
out_md = run_dir / "competitor_analysis.md"
|
||||
out_md.write_text(md, encoding="utf-8")
|
||||
return run_dir
|
||||
|
||||
@ -5,6 +5,7 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import re
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
@ -110,6 +111,113 @@ def generate_comment_sentiment_analysis_llm(payload: dict[str, Any]) -> str:
|
||||
return _call_llm(SENTIMENT_LLM_SYSTEM, user)
|
||||
|
||||
|
||||
def split_competitor_report_for_bridges(
|
||||
md: str, *, max_excerpt: int = 1200
|
||||
) -> dict[str, dict[str, str]]:
|
||||
"""
|
||||
按「## 一、」…「## 九、」切分规则报告，供大模型按章写衔接分析。
|
||||
每键含完整标题行与正文摘录（过长截断）。
|
||||
"""
|
||||
pat = re.compile(r"^## ([一二三四五六七八九])、([^\n]*)$", re.MULTILINE)
|
||||
matches = list(pat.finditer(md))
|
||||
out: dict[str, dict[str, str]] = {}
|
||||
for i, m in enumerate(matches):
|
||||
key = m.group(1)
|
||||
rest = m.group(2)
|
||||
title = f"## {key}、{rest}"
|
||||
start = m.end()
|
||||
end = matches[i + 1].start() if i + 1 < len(matches) else len(md)
|
||||
body = md[start:end].strip()
|
||||
exc = body[:max_excerpt]
|
||||
if len(body) > max_excerpt:
|
||||
exc += "\n\n…（本节摘录已截断）\n"
|
||||
out[key] = {"title": title, "excerpt": exc}
|
||||
return out
|
||||
|
||||
|
||||
def _parse_llm_json_object(text: str) -> dict[str, Any]:
|
||||
raw = (text or "").strip()
|
||||
if not raw:
|
||||
return {}
|
||||
if raw.startswith("```"):
|
||||
raw = re.sub(r"^```(?:json)?\s*", "", raw, flags=re.IGNORECASE)
|
||||
raw = re.sub(r"\s*```\s*$", "", raw)
|
||||
try:
|
||||
obj = json.loads(raw)
|
||||
return obj if isinstance(obj, dict) else {}
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
m = re.search(r"\{[\s\S]*\}", raw)
|
||||
if m:
|
||||
try:
|
||||
obj = json.loads(m.group(0))
|
||||
return obj if isinstance(obj, dict) else {}
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
return {}
|
||||
|
||||
|
||||
def _normalize_section_bridge_map(d: dict[str, Any]) -> dict[str, str]:
|
||||
allowed = frozenset("一二三四五六七八九")
|
||||
out: dict[str, str] = {}
|
||||
for k, v in d.items():
|
||||
if not isinstance(k, str) or len(k) != 1 or k not in allowed:
|
||||
continue
|
||||
if isinstance(v, str) and v.strip():
|
||||
out[k] = v.strip()
|
||||
return out
|
||||
|
||||
|
||||
BRIDGE_SECTIONS_SYSTEM = """你是竞品监测报告的**章节衔接**撰稿助手。
|
||||
|
||||
**输入 JSON** 含：
|
||||
- ``keyword``：监测词；
|
||||
- ``competitor_brief``：与本报告一致的**结构化摘要**（已裁剪体积）；
|
||||
- ``sections``：键为汉字「一」～「九」，每项含 ``title``（该章完整二级标题行）与 ``excerpt``（该章正文开头摘录，可能已截断）。
|
||||
|
||||
**任务**：为 **sections 中出现的每一键** 各写一段 **衔接性分析**（帮读者从摘要与摘录过渡到读该章表格/图），并与 ``competitor_brief`` 中的数字与结论一致。
|
||||
|
||||
**硬性要求**：
|
||||
- **仅输出一个 UTF-8 JSON 对象**（不要用 markdown 代码围栏包裹整段输出）；
|
||||
- 键必须为「一」「二」…「九」之一，且 **只对输入 sections 里存在的键** 给出字符串值；可省略无材料的键；
|
||||
- 每个值为 **Markdown 片段**（约 3～10 句中文），**禁止**使用 ``## `` 开头的行（不要写新的二级章标题）；可使用 ``###`` / ``####`` 或加粗小标题；
|
||||
- 所有**定量表述**须能在 ``competitor_brief`` 或对应 ``excerpt`` 中找到依据，**禁止编造** SKU 数、份额、价格；
|
||||
- 不要复述整章表格；不要写「详见下文矩阵」以外的空洞套话；可点出该章阅读重点（如价盘带、矩阵细类、评价规则局限等）。"""
|
||||
|
||||
|
||||
def generate_section_bridges_llm(
|
||||
*,
|
||||
keyword: str,
|
||||
brief: dict[str, Any],
|
||||
sections: dict[str, dict[str, str]],
|
||||
) -> dict[str, str]:
|
||||
"""一次 LLM 调用，返回各章衔接 Markdown 片段（键：一～九）。"""
|
||||
if not sections:
|
||||
return {}
|
||||
compact = compact_brief_for_llm(brief, max_chars=100_000)
|
||||
sec: dict[str, dict[str, str]] = {
|
||||
k: {"title": v.get("title", ""), "excerpt": v.get("excerpt", "")}
|
||||
for k, v in sections.items()
|
||||
if isinstance(v, dict)
|
||||
}
|
||||
for max_exc in (1200, 900, 600, 400, 280):
|
||||
for v in sec.values():
|
||||
ex = v.get("excerpt") or ""
|
||||
if len(ex) > max_exc:
|
||||
v["excerpt"] = ex[:max_exc] + "\n…\n"
|
||||
payload = {
|
||||
"keyword": keyword,
|
||||
"competitor_brief": compact,
|
||||
"sections": sec,
|
||||
}
|
||||
raw = json.dumps(payload, ensure_ascii=False)
|
||||
if len(raw) <= 92_000:
|
||||
break
|
||||
user = "请严格按系统说明，**只输出一个 JSON 对象**（键为一～九，值为 Markdown 字符串）：\n\n" + raw
|
||||
text = _call_llm(BRIDGE_SECTIONS_SYSTEM, user)
|
||||
return _normalize_section_bridge_map(_parse_llm_json_object(text))
|
||||
|
||||
|
||||
STRATEGY_SYSTEM = """你是市场策略顾问，根据**结构化监测摘要**与业务侧填写的**决策字段**，把「规则底稿」润色为可读的策略 Markdown。
|
||||
|
||||
**规则**：
|
||||
|
||||
@ -11,6 +11,7 @@ from .models import JdProduct, JdProductSnapshot, JobStatus, PipelineJob
|
||||
_REPORT_CONFIG_ALLOWED_KEYS = frozenset(
|
||||
{
|
||||
"llm_comment_sentiment",
|
||||
"llm_section_bridges",
|
||||
"comment_focus_words",
|
||||
"comment_scenario_groups",
|
||||
"external_market_table_rows",
|
||||
@ -29,6 +30,9 @@ def validate_report_config_body(value: dict) -> dict:
|
||||
if "llm_comment_sentiment" in value and value["llm_comment_sentiment"] is not None:
|
||||
if not isinstance(value["llm_comment_sentiment"], bool):
|
||||
raise serializers.ValidationError("llm_comment_sentiment 须为 true 或 false")
|
||||
if "llm_section_bridges" in value and value["llm_section_bridges"] is not None:
|
||||
if not isinstance(value["llm_section_bridges"], bool):
|
||||
raise serializers.ValidationError("llm_section_bridges 须为 true 或 false")
|
||||
raw = json.dumps(value, ensure_ascii=False)
|
||||
if len(raw) > 120_000:
|
||||
raise serializers.ValidationError("报告配置体积过大")
|
||||
@ -41,6 +45,7 @@ _ARTIFACT_FILES: tuple[tuple[str, str], ...] = (
|
||||
("comments", "comments_flat.csv"),
|
||||
("detail_ware", "detail_ware_export.csv"),
|
||||
("report", "competitor_analysis.md"),
|
||||
("section_bridge_llm", "section_bridge_llm.json"),
|
||||
)
|
||||
|
||||
|
||||
|
||||
@ -20,11 +20,15 @@ export function useReportConfigForm() {
|
||||
const marketRows = ref([
|
||||
{ indicator: '', value_and_scope: '', source: '', year: '' },
|
||||
])
|
||||
const useLlmCommentSentiment = ref(false)
|
||||
const useLlmSectionBridges = ref(false)
|
||||
|
||||
function resetToEmpty() {
|
||||
focusWordRows.value = [{ text: '' }]
|
||||
scenarioGroups.value = [{ label: '', triggersText: '' }]
|
||||
marketRows.value = [{ indicator: '', value_and_scope: '', source: '', year: '' }]
|
||||
useLlmCommentSentiment.value = false
|
||||
useLlmSectionBridges.value = false
|
||||
}
|
||||
|
||||
/**
|
||||
@ -94,6 +98,9 @@ export function useReportConfigForm() {
|
||||
} else {
|
||||
marketRows.value = [{ indicator: '', value_and_scope: '', source: '', year: '' }]
|
||||
}
|
||||
|
||||
useLlmCommentSentiment.value = Boolean(cfg.llm_comment_sentiment)
|
||||
useLlmSectionBridges.value = Boolean(cfg.llm_section_bridges)
|
||||
}
|
||||
|
||||
/** @returns {Record<string, unknown>} 可 PATCH 到后端的 report_config；全空则为 {} */
|
||||
@ -133,6 +140,9 @@ export function useReportConfigForm() {
|
||||
}))
|
||||
}
|
||||
|
||||
out.llm_comment_sentiment = useLlmCommentSentiment.value
|
||||
out.llm_section_bridges = useLlmSectionBridges.value
|
||||
|
||||
return out
|
||||
}
|
||||
|
||||
@ -178,6 +188,8 @@ export function useReportConfigForm() {
|
||||
focusWordRows,
|
||||
scenarioGroups,
|
||||
marketRows,
|
||||
useLlmCommentSentiment,
|
||||
useLlmSectionBridges,
|
||||
resetToEmpty,
|
||||
applyFromApiConfig,
|
||||
buildPayload,
|
||||
|
||||
@ -31,6 +31,8 @@ const {
|
||||
focusWordRows,
|
||||
scenarioGroups,
|
||||
marketRows,
|
||||
useLlmCommentSentiment,
|
||||
useLlmSectionBridges,
|
||||
applyFromApiConfig,
|
||||
buildPayload,
|
||||
addFocusRow,
|
||||
@ -247,8 +249,18 @@ watch(
|
||||
<div v-if="selectedId" class="report-config-block">
|
||||
<h3 class="report-config-title">报告里的评价统计怎么算</h3>
|
||||
<p class="hint-top report-config-hint">
|
||||
下面三项都<strong>可以不改</strong>：留空并保存，表示沿用系统内置规则。请先点「保存以上设置」，再点「重新生成报告」（需要大模型时先勾选页面上方对应选项）。
|
||||
下面几项都<strong>可以不改</strong>：留空并保存，表示沿用系统内置规则。请先点「保存以上设置」，再点「重新生成报告」；需要大模型时勾选下方「评价归纳」或「各章衔接」，页顶「使用大模型生成」仍用于整份报告另一种生成模式。
|
||||
</p>
|
||||
<div class="toolbar toolbar-llm-flags">
|
||||
<label class="chk-inline">
|
||||
<input v-model="useLlmCommentSentiment" type="checkbox" />
|
||||
评价章大模型归纳（§8.2 主题解读，需 API 密钥）
|
||||
</label>
|
||||
<label class="chk-inline">
|
||||
<input v-model="useLlmSectionBridges" type="checkbox" />
|
||||
各章大模型衔接分析（第一～九章标题后各一段，一次调用多章，需密钥）
|
||||
</label>
|
||||
</div>
|
||||
<div class="report-config-actions">
|
||||
<button
|
||||
type="button"
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user