fd-cn-report

MCP server for Chinese financial reports — 31 申万 L1 industry AI rule system, outline extraction, AI structured extraction, Elasticsearch store + search, and interactive rules dashboard.

Industry Rules System

21,698 LLM rules covering 31 申万 L1 industries with per-section extraction instructions generated from real annual report PDFs. Each rule maps an indicator to its exact section position in the periodic report.

Coverage

章节	内容	说明
资产负债表	全部科目 (112+ rules)	从合并资产负债表提取
利润表	全部科目 (130+ rules)	从合并利润表提取
现金流量表	全部科目 (42+ rules)	从合并现金流量表提取
管理层讨论	主营业务分析、成本、产销量、研发	从实际报告第三节提取
主要财务指标	ROE、EPS、毛利率、分季度数据	从第二节提取
报表附注/重要事项	关联交易、担保、诉讼、资产减值	从第六节+第十节附注提取
股东信息	前十大股东、持股变动、分红	从第七节提取
公司治理	董事会、高管薪酬	从第四节提取
环境	排污、碳排放、能源消耗	从第五节提取

Industry-Specific Rules

6 industries have industry-specific rules on top of the universal set:

行业	特有指标	来源
银行	不良率、净息差、资本充足率、贷款五级分类	工商银行年报
房地产	合同负债、存货-开发成本、土储、销售面积	保利发展年报
电力设备	在建工程、产能利用率、应收账款	宁德时代年报
医药生物	研发费用、销售费用、在研管线、无形资产	恒瑞医药年报
非银金融	保费收入、赔付支出、偿付能力、新业务价值	中国平安年报
农林牧渔	存货-消耗性生物资产	隆平高科年报

Rule Generation Pipeline

Rules are generated from real annual report PDFs via parallel LLM calls:

# Generate rules for all 31 industries from real reports
python scripts/generate_rules_from_real_reports.py --max-concurrent 5

# Single industry
python scripts/generate_rules_from_real_reports.py --industry 801120

# Run industry-specific sections only
python scripts/generate_rules_from_real_reports.py --industry 801780 --llm-only

Dashboard

Interactive web UI to browse, filter, and search all 21,698 rules:

# CLI
python scripts/industry_rules_dashboard.py
# or
fd-cn-report

# MCP tool
open_industry_rules_dashboard(port=8888)

Features:

• Industry filter — dropdown with all 31 申万 L1 industries
• Module filter — filter by balance_sheet, income_statement, etc.
• Keyword search — indicator name, section, instruction text
• Sortable columns — click header to sort
• Pagination — 50 rules per page
• Color-coded tags — module type visualization
• Real-time stats — filtered count, industry count, module count

---

Tools (44 MCP tools)

Layer	Tool	Description
Company API	get_company	Resolve ticker/name → company entry
	list_filings	List CNINFO disclosures by form/category + year
	get_filing	One announcement's metadata + PDF URL
	get_financials	Income/balance/cashflow via akshare
	get_financial_statements	三大报表 as text from the annual-report PDF
	get_section	(ticker, year, section) → section text
	list_report_types	Browse CNINFO disclosure category catalog
	get_special_report	Special-type report (招股说明书, 收购报告书, …)
HK Stock	get_hk_company	Resolve HK stock by ticker/name
	list_hk_filings	List HKEX filings
	get_hk_financials	HK financial statements
	get_hk_section	HK report section extraction
Official-website (SSE/SZSE/BSE)	get_sse_company	Resolve SSE-listed company by 6-digit ticker
	list_sse_filings	List SSE disclosures from sse.com.cn
	get_sse_section	SSE annual-report section extraction
	get_sse_interaction	上证e互动 investor Q&A
	get_szse_company	Resolve SZSE-listed company by 6-digit ticker
	list_szse_filings	List SZSE disclosures from szse.cn
	get_szse_section	SZSE annual-report section extraction
	get_szse_interaction	互动易 investor Q&A
	get_bse_company	Resolve BSE-listed company by 6-digit ticker
	list_bse_filings	List BSE disclosures (BSE-native, CNINFO fallback)
	get_bse_section	BSE annual-report section extraction
Official-website (CSRC)	list_csrc_filings	List CSRC regulatory announcements
	get_csrc_ipo_review	CSRC IPO (首发) review status
	get_csrc_merger_review	CSRC 并购重组 review status
	list_csrc_enforcement	CSRC enforcement actions
Ministry statistics	list_ministries	List supported ministry stat sources
	get_ministry_stat	Ministry stats page -> HTML tables
	get_nbs_stat	NBS macro statistic by indicator code
PDF / AI / ES	list_outline	Parse 目录 from report URL or PDF path
	extract_section	Body text by exact title / regex / ordinal
	ai_extract	LLM-structured extraction over section text
	index_records	Bulk index records into ES
	search_reports	BM25 + filter search with highlights
	delete_index	Drop cnreport-{year} index
Cache	list_cache	List cached reports
	clear_cache	Evict cached reports
Indicators	list_indicators	Browse indicator rule set
	get_indicator	One indicator's value
	extract_indicators	All applicable indicators in one pass
	extract_indicators_by_position	CSV-driven extraction
	extract_indicators_batch	Batch concurrent extraction
Dashboard	open_industry_rules_dashboard	Start the rules web dashboard

Typical Chain

# 1. Resolve company → 2. find latest annual → 3. pull MD&A → 4. LLM-extract revenue table

co = get_company("600519")
# {"stock_code": "600519", "name": "贵州茅台", "org_id": "gssh0600519", "exchange": "sse", ...}

filings = list_filings("600519", form="年度报告", year=2023, limit=3)
# [{"announcement_id": "1219730876", "pdf_url": "http://static.cninfo.com.cn/.../*.PDF", ...}]

sec = get_section("600519", year=2023, section="管理层讨论与分析")
# {"text": "<full MD&A body>", "pdf_url": "...", "outline_entry": {...}, ...}

records = ai_extract(
    text=sec["text"],
    schema={"type": "object", "properties": {
        "segment": {"type": "string"},
        "revenue_2023": {"type": "string"},
    }, "required": ["segment", "revenue_2023"]},
)
# {"records": [{"segment": "茅台酒", "revenue_2023": "139,989,000,000"}, ...]}

Special Report Types

CNINFO exposes dozens of disclosure categories beyond the four periodic reports (招股说明书, 增发, 业绩预告, 收购报告书, 股权激励, …). Browse the catalog, then list or retrieve by category:

catalog = list_report_types()
# {"groups": [{"name": "定期报告", "categories": [...]}, {"name": "融资", ...}, ...], "count": 26}

list_report_types(group="融资")
# {"group": "融资", "categories": [{name: "首发", code: "category_sf_szsh", ...}, ...], "count": 6}

filings = list_filings("600519", category="首发", limit=3)

sec = get_special_report("600519", category="首发", section="募集资金运用")

三大报表 (Three Major Financial Statements)

get_financials returns akshare's structured numeric tables. get_financial_statements pulls the three major statement sections as text from the PDF:

stmts = get_financial_statements("600519", year=2023)
# {
#   "stock_code": "600519", "company_name": "贵州茅台", "year": 2023,
#   "form": "年度报告", "pdf_url": "...", "cached": False,
#   "statements": {
#     "income_statement": {"title": "2、 合并利润表", "outline_entry": {...}, "char_count": 4521, "text": "..."},
#     "balance_sheet":    {"title": "1、 合并资产负债表", ...},
#     "cashflow":         {"title": "3、 合并现金流量表", ...},
#   },
#   "missing": [],
# }

Report Cache

Every report fetch goes through an on-disk cache under .cache/reports/. First fetch downloads PDF + extracts text + outline; subsequent fetches read from disk.

list_cache()
# {"cache_dir": ".../.cache/reports", "count": 2, "entries": [...]}

clear_cache()                              # evict everything
clear_cache(stock_code="600519")           # evict one company
clear_cache(stock_code="600519", year=2023) # evict one company + year

Indicators

The indicator engine profiles each company, filters applicable rules, and routes each indicator to akshare / report-section / computed / external.

# Preview → pull one → pull all → CSV
list_indicators(company="工商银行")           # rules applicable to 工商银行
get_indicator("资本充足率", "工商银行", 2023)  # one value
extract_indicators("工商银行", 2023)           # all applicable, one PDF fetch
extract_indicators("工商银行", 2023,
                   indicators=["资本充足率","不良率"])  # subset
extract_indicators("工商银行", 2023, extractor_mode="python")  # LLM-free
extract_indicators_by_position("工商银行", 2023)  # CSV-driven

Multi-form support (年度报告 / 半年度报告 / 第一季度报告 / 第三季度报告):

extract_indicators_by_position("工商银行", 2023, form="第一季度报告")
extract_indicators("贵州茅台", 2023, form="半年度报告")

Concurrency

extract_indicators("工商银行", 2023, concurrency=4)  # explicit cap
extract_indicators("工商银行", 2023, concurrency=1)  # sequential

# Batch: multi-company concurrent extraction
extract_indicators_batch([("601398", 2023), ("600519", 2023)],
                         concurrency=2, extract_concurrency=4)
# → {"results": {"601398_2023": {...}, ...}, "failures": [...], "concurrency": 2}

Section Cache

LLM responses are persisted to disk, keyed by (pdf_url, section_key, period, rules_hash). Subsequent runs reuse cached records. Set LLM_SECTION_CACHE=off to disable.

Standalone CLIs

# Full engine extraction
python scripts/extract_indicators.py 601398 --year 2023 \
    [--rules indicator_rules.json] [--extractor auto|llm|python] \
    [--indicators 资本充足率,不良率] [--out-dir ./out]

# CSV-driven extraction
python scripts/extract_indicators_by_position.py 601398 --year 2023 \
    [--csv docs/indicators_position.csv] [--extractor auto|llm|python] \
    [--form 年度报告|半年度报告|第一季度报告|第三季度报告]

# Multi-year extraction
python scripts/extract_indicators_multiyear.py 601398 2023 2024

# Industry rules dashboard
python scripts/industry_rules_dashboard.py [port]
# or: fd-cn-report

# Generate industry rules from real reports
python scripts/generate_rules_from_real_reports.py

# Check industry coverage
python scripts/check_industry_coverage.py

# Seed industry rules
python scripts/seed_industry_rules.py

HK Stock Support

get_hk_company("00700")                    # → 腾讯控股
list_hk_filings("00700", year=2023)        # → HKEX filings
get_hk_financials("00700")                 # → financial statements
get_hk_section("00700", year=2023, section="管理层讨论与分析")

Official-website datasources (SSE / SZSE / BSE)

Direct, primary-source disclosure paths complementing CNINFO. Each exchange client resolves a 6-digit ticker locally (no network) and lists disclosures from the exchange's own site. BSE falls back to CNINFO when its own API is thin, tagging each row with source: "bse" | "cninfo". Section extraction reuses the same outline pipeline as CNINFO/HK.

# SSE (上交所) - 600/601/603/605/688/900 codes
get_sse_company("600519")
list_sse_filings("600519", year=2023)                         # -> sse.com.cn disclosures
get_sse_section("600519", year=2023, section="管理层讨论与分析")
get_sse_interaction("600519")                                  # -> 上证e互动 Q&A

# SZSE (深交所) - 000/001/002/003/300/301 codes
get_szse_company("000001")
list_szse_filings("000001", year=2023)                        # -> szse.cn disclosures
get_szse_section("000001", year=2023, section="管理层讨论与分析")
get_szse_interaction("000001")                                 # -> 互动易 Q&A (irm.cninfo.com.cn)

# BSE (北交所) - 430xxx / 83xxxx / 87xxxx / 88xxxx / 920xxx codes
get_bse_company("835185")
list_bse_filings("835185", year=2023)                         # -> BSE-native, else CNINFO fallback
get_bse_section("835185", year=2023, section="管理层讨论与分析")  # result carries `source`

Endpoints are undocumented. SSE (query.sse.com.cn), SZSE (www.szse.cn/api), BSE (www.bse.cn), and the Q&A hosts (sns.sseinfo.com, irm.cninfo.com.cn) expose no official API contract and shift over time. Clients are thin, retry 429/5xx, and bypass proxy env (trust_env=False). Run CNREPORT_SELFCHECK_LIVE=1 uv run python selfcheck.py to ping each endpoint - 4xx/5xx are flagged per source without failing the suite. Name-fragment resolution is not supported by these clients; resolve the name via get_company (CNINFO) first, then pass the 6-digit code.

CSRC regulatory data (证监会)

CSRC regulatory data CNINFO/exchanges do not cover: announcements, IPO / 并购重组 review status, and enforcement. HTML-parsed with lxml; endpoints undocumented.

list_csrc_filings(begin_date="2024-01-01")              # -> regulatory announcements
get_csrc_ipo_review("贵州茅台")                          # -> IPO review status row
get_csrc_merger_review("某科技股份公司")                 # -> M&A review status row
list_csrc_enforcement()                                 # -> administrative-penalty actions

Ministry statistics (部级部门)

Structured economic/financial statistics from ministry-level departments, complementing fd-cn-gov's catalog-archive scraping with data queries. NBS has a JSON API; the others publish HTML tables. Results are TTL-cached under .cache/stats/. Base URLs reuse fd-cn-gov's registry when importable.

list_ministries()                       # -> [{id, label, en, transport, base}, ...]
get_nbs_stat("A0201")                   # -> NBS GDP series {period: value}  (dbcode="hgnd" annual)
get_ministry_stat("gacc")               # -> GACC trade page parsed into HTML tables
get_ministry_stat("pboc", limit=20)     # -> PBoC monetary tables
# Supported ids: nbs mof pboc safe gacc nfra

Ministry stat-page paths and CSRC URLs are best-guess and undocumented - verify live via CNREPORT_SELFCHECK_LIVE=1 uv run python selfcheck.py. Corrections are one-line edits in ministry_stats_client._MINISTRIES / csrc_client._URLS.

Setup

uv sync                    # installs akshare, pypdf, fastmcp, ...
uv run python server.py    # FastMCP over stdio

Self-check (no network):

uv run python selfcheck.py           # DB + outline + company API + special reports
uv run python selfcheck_cache.py     # report cache + three-statements extraction

Tests (offline):

uv run --with pytest python -m pytest test_cnreport.py -v -p no:logfire

Configuration

CNINFO and akshare are keyless. Other tools need env vars in .env:

Var	Used by	Required?
LLM_API_KEY, LLM_BASE_URL, LLM_MODEL	ai_extract, rule generation	Yes for AI
ES_URL (+ optional ES_API_KEY or ES_USERNAME/ES_PASSWORD)	index_records, search_reports, delete_index	Yes for ES
DAAS_DATABASE_URL	provenance writes, rules storage	Defaults to daas.db
CNREPORT_CACHE_DIR	report cache	Defaults to .cache/reports/

Architecture

fd-cn-report/
├── server.py                      # FastMCP server (@app.tool registrations)
├── cninfo_client.py               # CNINFO API (keyless lookup + query)
├── hk_stock_client.py             # HK stock API (akshare + HKEX)
├── financials_client.py           # akshare financial statements (lazy)
├── cnreport_tools.py              # Pure helpers (outline, extract, LLM, ES)
├── report_cache.py                # On-disk PDF + text + outline cache
├── llm_section_cache.py           # LLM response section cache
│
├── indicators_client.py           # Rules engine (load, profile, route, extract)
├── indicators_extractors.py       # Pluggable Python extractors
├── indicators_models.py           # Pydantic extraction models
├── indicators_csv_migration.py    # CSV → JSON rule migration
├── report_section_map.py          # Section alias expansion + matching
│
├── rules_db.py                    # Rules database (SQLite via SQLAlchemy)
├── rules_models.py                # Pydantic rule models
├── rules_skills.py                # LLM rule generation + validation
├── cnreport_models.py             # ORM models (LlmRule, ScriptRule)
├── cnreport_database.py           # Database connection management
│
├── industry_taxonomy.py           # 申万 L1 industry taxonomy
├── industry_coverage.py           # Coverage checks for industry rules
│
├── docs/
│   ├── industry_taxonomy.json     # 31 industry taxonomy
│   ├── industry_indicator_baseline.json  # Baseline indicators per industry
│   ├── indicators_position.csv    # Indicator catalog (CSV source)
│   └── indicators-methodology.md  # Rendered methodology
│
├── scripts/
│   ├── industry_rules_dashboard.py          # Web dashboard (CLI + MCP)
│   ├── generate_rules_from_real_reports.py  # Per-industry rule generation
│   ├── generate_all_industry_rules.py       # LLM rule generation
│   ├── seed_industry_rules.py               # Universal rule seeding
│   ├── seed_missing_industry_rules.py       # Industry-specific seeding
│   ├── extract_indicators.py                # Standalone extraction CLI
│   ├── extract_indicators_by_position.py    # CSV-driven extraction CLI
│   ├── extract_indicators_multiyear.py      # Multi-year batch extraction
│   ├── check_industry_coverage.py           # Coverage validation
│   ├── migrate_indicators_csv.py            # CSV → DB migration
│   └── rules_dashboard.py                   # indicator_rules.json editor
│
└── .cache/reports/                # Downloaded PDFs + extracted text + outlines

Adding a New Industry

1. Add to docs/industry_taxonomy.json:

   {"industry": "801xxx", "label": "行业名称", "report_kinds": ["annual-report", "interim-report", "quarterly-report"]}
   

2. Add baseline indicators to docs/industry_indicator_baseline.json:

   {"cn/801xxx/listed/annual-report": ["资产总计", "营业收入", "净利润", ...]}
   

3. Add representative company to scripts/generate_rules_from_real_reports.py:

   "801xxx": ("600xxx", "代表公司", "行业名称"),
   

4. Define industry-specific sections:

   "801xxx": [("section_name", "keyword", "keyword2", "guidance"), ...],
   

5. Generate rules:

   python scripts/generate_rules_from_real_reports.py --industry 801xxx
   

License

MIT