credit-risk-extraction

Structured Credit Risk Extraction

Target Role

Role: Risk Analysis Expert / Credit Approval Officer / Risk Control Rule Datafication Expert
Scenario: Pre-loan due diligence report structuring, approval committee material preparation, post-loan risk ledger maintenance, risk report quality review, industry risk knowledge base construction
Output Purpose: Generate an enterprise label profile and structured risk analysis report, transforming abstract risks into specific data-verifiable rules
Decision Level: Decision support (output is for reference by credit review personnel and is not directly used for approval decisions)
Execution Frequency: Executed on demand, usually invoked during pre-loan due diligence or periodic risk reviews

Data Sources

Required Data

Data Item	Source	Acquisition Method	Sensitivity Level
Enterprise credit report/risk evaluation text	User upload	`file read`	Internal
Enterprise basic situation description	User upload	`file read`	Internal
GB/T 4754-2017 national standard industry classification	Context/reference materials	`context`	Public
Public information sources (business registration/credit reporting/public opinion)	External API/web search	`api call` / `web search`	Public

Data Desensitization Rules

Customer name: Replace with [客户名称]
Bank account number: Replace with [银行账号]
Personal ID number: Replace with [证件号码]
Contact phone number: Replace with [联系电话]
Detailed address: Retain province, city, and district; replace the detailed address with [详细地址]

Degradation Strategy

If the enterprise credit report is missing: output the error message "Missing required input: enterprise credit report/risk evaluation text" and refuse to execute
If the enterprise basic situation description is missing: require the user to provide it, or execute only risk point extraction (skip label profile filling)
If public information sources are unavailable: execute extraction based only on the text provided by the user, and annotate "not verified using external data sources"
If the GB/T 4754-2017 standard is unavailable: use the built-in industry classification reference (see references/tag-system.md), and annotate in the output "industry classification is based on the built-in reference; verification is recommended"
If the number of risk points in the original text is unclear: perform automatic counting, and state in the output "the number of risk points is automatically counted by the system; user verification is recommended"

Workflow

📋 Strictly follow the 12 standards defined in METHODOLOGY.md.

Step 0: Data confirmation (read before write) (Data source:`user_upload`, Executing entity:`ai`, Confirmation mechanism:`none`）

List the input files: corporate credit report/risk assessment text, basic corporate information description
Confirm whether the original text contains sections such as "Key Risk Assessment of Credit Business," "Matters Recommended for Attention," and "Risk Warnings"
If the number of risk points is provided by the user, record it as N; if not provided, perform automatic counting
Proceed to subsequent steps only after validation passes: If the corporate credit report is missing, output an error message and terminate

Step 1: Information extraction and risk point counting (Data source:`user_upload`, Execution entity:`ai`, Confirmation mechanism:`none`）

Locate sections such as "Key Risk Assessment of Credit Business," "Matters Recommended for Attention," and "Risk Warnings," and scan paragraph by paragraph
Number each independent risk point (such as R1, R2, R3…), and record the original wording
Risk point count locked: After extraction is completed, confirm N risk points; subsequent output must cover all N, with none omitted
If the risk description is nested (one paragraph contains multiple risks), it must be broken down one by one and counted as independent risk points
Anti-laziness instruction: Do not skip any risk point; all risk points must be extracted one by one

Step 2: Precise industry classification positioning (Data source:`context` + `references/tag-system.md`, Execution entity:`ai`, Confirmation mechanism:`none`）

According to the GB/T 4754-2017 national standard industry classification standard, perform a four-level precise classification of the target enterprise
Path display order: Section → Division → Group → Class, and each level must be labeled with its code and name
If the enterprise involves multiple main businesses, determine the main classification according to the business with the largest revenue contribution, and note the remaining businesses as "sideline"
It is prohibited to stop after matching only to the division or group, and it must penetrate through to the class (four-digit code)
Consult references/tag-system.md to obtain the industry classification standard format

Step 3: Enterprise tag profile filling (Data source:`user_upload` + `references/tag-system.md`, Execution entity:`ai`, Confirmation mechanism:`none`）

Referring to the tag system reference (consult references/tag-system.md), fill in the enterprise tag profile category by category
A total of 14 categories of tags—basic attributes (5 categories), operating characteristics (4 categories), risk characteristics (3 categories), and external environment (2 categories)—must be checked item by item
For those with clear basis: directly fill in the tag value, and a brief basis explanation may be attached
For tags that cannot be defined from the original text: fill in "(not mentioned in the original text)"; do not fill in by speculation
Capital attribute and group attribute are required fields, and if not mentioned in the original text, it must be stated as "(needs verification)"
Anti-laziness instruction: Do not skip any tag category; each item must be checked and filled in

Step 4: Structured transformation of risk points (Data source:`user_upload` + `references/scenario-rules.md`, Execution entity:`ai`, Confirmation mechanism:`none`）

Transform each risk point (R1…RN) one by one according to the structured format; do not merge or skip any risk point
Risk definition: Use one sentence to precisely describe the essence of the risk, without grading (the use of "high/medium/low risk" is prohibited), focusing on "under what circumstances what consequences will occur"
Original text description: Copy exactly consistent with the original text; reduction, rewriting, or summarization is strictly prohibited
Risk analysis direction: Break down into 2-4 specific analysis dimensions, and each dimension must state "what data to use, through what method, and what fact to verify"
Data source: Distinguish between known data sources (already provided in the source text) and inferable data sources (information channels requiring additional acquisition)
Applicable scenarios: Based on four types of data sources—business registration/credit reporting/annual reports/public sentiment—express in the format of "data field + judgment condition" (refer to references/scenario-rules.md）
Suggested points of attention: Provide specific investigation dimensions, practical scripts, or query channels; specific indicator thresholds are prohibited (refer to references/scenario-rules.md）
Anti-laziness instruction: Each risk point must cover 6 dimensions; no dimension may be skipped

Step 5: Completeness self-check (data source:`context`, executing entity:`ai`, confirmation mechanism:`none`）

Check item by item the consistency between the extracted risk points and the quantity locked in Step 1
Self-check list：
- Number of risk points: the number of output risk points = the N counted in Step 1, with no omissions
- Source text source text: the "enterprise risk source text description" of each risk point has been copied completely, without reduction
- Complete labels: capital attributes and group attributes have been filled in (even if "to be verified")
- No grading vocabulary: the full text does not contain "high risk", "medium risk", or "low risk"
- No vague vocabulary: the full text does not contain "related", "certain", "appropriate", "as much as possible", or "in principle"

Step 6: Format integration output (data source:`context` + `assets/risk-extraction-report-template.md`, executing entity:`ai`, confirmation mechanism:`none`）

Organize the two sections 【Enterprise Label Profile】 and 【Structured Risk Analysis】 according to the preset format, ensuring standardized formatting and clear hierarchy
Label profile first, risk analysis after
Each risk point is marked with a level-three heading (###), and the heading is the risk category name
Risk points are arranged in the order they appear in the source text and must not be reordered
Finally attach a brief completeness statement: "This analysis has covered all N risk points in the source text"
Include a disclaimer (cite assets/risk-extraction-report-template.md）

Output Format (Output Format)

The output is a complete Markdown structured risk extraction report, template reference assets/risk-extraction-report-template.md。

【Enterprise Label Profile】

- **行业分类**：【门类代码】【门类名称】>【大类代码】【大类名称】>【中类代码】【中类名称】>【小类代码】【小类名称】
- **基础属性**：【规模】、【所有制性质】、【资本属性】、【组织架构】、【生命周期】
- **经营特征**：【供应链特征】、【生产模式】、【技术特征】、【销售模式】
- **风险特征**：【合规风险】、【财务风险】、【司法风险】
- **外部环境**：【政策环境标签】、【市场环境标签】

【Structured Risk Analysis】

Output independently for each risk point:

### [风险类别名称，如"供应链安全风险"]
**风险定义**：【一句话精准描述风险本质，不含分级】
**企业风险原文描述**：【与原文完全一致，严禁缩减】
**风险分析方向**：[具体分析维度和逻辑]
**数据来源说明**：[已知数据源] + [可推断数据源]
**适用场景**：[基于工商/征信/年报/舆情四类数据源，体现"数据字段+判定条件"]
**建议关注事项**：[具体调查维度、话术、渠道，禁止提供具体指标阈值]

Report structure：

Basic report information (enterprise name, extraction time, applicable industry, list of input materials)
【Enterprise Label Profile】 (industry classification down to the fourth level, basic attributes, operating characteristics, risk characteristics, external environment)
【Structured Risk Analysis】 (statement of the total number of risk points + 6-dimension structured analysis of each risk point)
Completeness statement ("This analysis has covered all N risk points in the source text, with no omissions")
Data sources and disclaimer (must include "does not constitute a credit approval opinion" and "for reference only")

This output can be parsed and used by credit-industry-analysis and visit-memo。

Compliance Constraints (Constraints)

Risk grading prohibited: Grading expressions such as "high risk", "medium risk", and "low risk" are strictly prohibited; only objectively describe risk facts
Refined industry classification: GB/T 4754-2017 must be used, with display order: section→division→group→class, and must not stop at division or group
Complete label system: Extended labels such as capital attributes and group attributes must be included, and labels must have a basis
Vague vocabulary prohibited: Strictly prohibited to use "related", "certain", "appropriate", "as much as possible", "in principle"
Content completeness: If the original text has N risk points, the output must cover all N points; omissions will lead to serious consequences
Prohibition of return commitments: The output must not contain any deterministic return commitments or investment advice
Data timeliness annotation: If external data sources are used (such as business registration/credit data), the data query date must be annotated
Prohibition of exceeding authority advice: Only risk identification and structured advice may be provided; it must not cross into areas beyond authority such as credit approval or credit decision-making

Audit Trail

Generate an audit log after each execution:

{
  "skill_name": "credit-risk-extraction",
  "skill_version": "1.0.0",
  "execution_time": "2026-05-05T14:00:00+08:00",
  "company_name": "[客户名称]",
  "operator": "ai",
  "steps": [
    {
      "step": "数据确认",
      "executor": "ai",
      "data_source": {"type": "user_upload"},
      "result": "pass"
    },
    {
      "step": "风险点提取",
      "executor": "ai",
      "data_source": {"type": "user_upload"},
      "risk_points_count": 3,
      "result": "pass"
    }
  ],
  "compliance_check": {
    "no_risk_grading": true,
    "no_vague_words": true,
    "industry_classification_4_levels": true,
    "all_risk_points_covered": true,
    "disclaimer_present": true
  },
  "audit_log_retention": "3年"
}

Gotchas

#1: Risk point merging causes omissions

Symptom: A paragraph in the original text contains multiple independent risks (such as supplier concentration + customer concentration), and the AI merges them into one risk point in the output.
Cause: The risk point splitting rule was not executed, and nested risks were treated as a single risk.
Solution: In Step 1, clearly state "if the risk description contains nesting, it must be broken down one by one and counted separately as independent risk points", and verify the count in the Step 5 self-check list.

#2: Rewriting the original risk description

Symptom: The output "original enterprise risk description" is inconsistent with the original text, with reduction, summarization, or rewriting.
Cause: The AI defaults to performing a text summarization task and does not strictly follow the "verbatim copy of the original text" principle.
Solution: In Constraints, reinforce "original description: copied exactly as the original text, strictly prohibited to reduce, rewrite, or summarize", and add an original text consistency check to the Step 5 self-check list.

#3: Industry classification not drilled down to the small class

Symptom: Industry classification is only displayed to the category or major class (such as "C Manufacturing" or "C36 Automobile Manufacturing"), without drilling down to the small class.
Cause: The four-level classification requirement was not strictly implemented, or the GB/T 4754-2017 standard reference is missing.
Solution: In Constraints, clearly state "GB/T 4754-2017 must be adopted, display order: category → major class → medium class → small class", and explicitly reference it in Step 2 references/tag-system.md。

#4: Capital attribute and organizational structure labels left blank

Symptom: When the original text does not mention capital attributes or organizational structure, the AI directly skips the label or leaves it blank.
Cause: Capital attributes and organizational structure were not identified as required fields, and the rule "if not mentioned in the original text, fill in '(needs verification)'" was not executed.
Solution: In Step 3, clearly state "capital attributes and group attributes are required fields; if not mentioned in the original text, state '(needs verification)'", and add a required fields check to the Step 5 self-check list.

Examples

Example 1: Risk extraction for a standard manufacturing enterprise

User input：

请对以下信贷报告进行风险结构化提取：

企业基本情况：XX汽车零部件制造有限公司，成立于2015年，主营汽车零部件制造，员工200人，年营收1.2亿元。

信贷报告风险评价：
1. 公司前三大供应商采购占比超过60%，且均为单一来源供应商，若任一供应商出现停产或质量问题，将直接影响公司正常生产。
2. 公司应收账款周转天数从2025年的90天延长至2026年Q1的150天，主要客户回款速度明显放缓。
3. 公司属于汽车零部件铸造环节，涉及废气废水排放，当地环保局近期提高了排放标准，公司需进行环保设施升级改造。

Skill execution flow：

Step 0: Confirm the input files (credit report, basic situation description), number of risk points N=3.
Step 1: Extract 3 risk points (R1 supply chain concentration, R2 extension of accounts receivable payment period, R3 environmental compliance), lock the count at N=3.
Step 2: Four-level industry classification positioning → C Manufacturing > C36 Automobile Manufacturing > C367 Automobile Parts and Accessories Manufacturing > C3670 Automobile Parts and Accessories Manufacturing.
Step 3: Label profile filling → basic attributes (medium-sized/private/non-listed/single/mature stage), operating characteristics (domestic procurement/discrete manufacturing/traditional manufacturing/direct sales), risk characteristics (need to pay attention to production safety/abundant cash flow/no judicial), external environment (policy neutral/stock competition).
Step 4: Structured transformation of risk points → each risk point is expanded across 6 dimensions (definition/original text/analysis direction/data source/applicable scenario/suggested matters to note).
Step 5: Completeness self-check → number of risk points 3=3, original text fully copied, no grading vocabulary, no vague vocabulary.
Step 6: Format integration output → generate a complete report (label profile + structured risk analysis + completeness statement + disclaimer).

Output key points：

Core conclusion: all 3 risk points in the original text have been covered, the industry classification goes all the way down to the fourth level, and the label profile is complete.
Key characteristics: no risk grading expressions, the original descriptions are fully copied, and each risk point includes 6-dimension structured analysis.
Deliverable: a complete Markdown structured risk extraction report (about 80 lines).

Example 2: Multi-risk point extraction for a group enterprise