Energy Expert:Guide to creating information extraction templates

1. Template types

AI Doc provides two types of information extraction templates to meet the requirements in various scenarios.

2. Prompt template creation guide

Scenarios

Prompt templates (guiding models to output structured/unstructured content through natural language instructions) are suitable for various task scenarios, especially excelling in tasks requiring flexibility, contextual understanding, free generation, or complex reasoning.

Prompt writing techniques

Direct questioning method

Scenarios

Suitable for scenarios with clear objectives, simple questions, and definite answers.

Writing key points

• Concise: Express questions in the shortest way possible. Overly lengthy questions may contain redundant information, causing the model to misunderstand or answer incorrectly.

  • Before optimization: "Our company recently received an invoice from a supplier, and we need you to extract some key information from it. Please carefully read this invoice document and tell me the supplier's company name, invoice number, and issue date. Additionally, if there is a total amount due on the invoice, please extract that as well. We need this information for accounting processing and payment arrangements."

  • After optimization: "Extract from supplier invoice: company name, invoice number, issue date, total amount due."

• Specific: In information extraction scenarios, clearly list all items to be extracted and provide detailed characteristic descriptions for each item.

  • Before optimization: "Analyze this article and extract key information."

  • After optimization: "Extract the following information from this article: the author's full name and any relevant titles (such as 'Professor', 'Researcher', etc.), the author's views on global warming, and 3-5 keywords."

• Avoid ambiguity: If a word or phrase may have multiple meanings, either clarify its meaning or rephrase to eliminate ambiguity.

  • Before optimization: "What is a tree?"

  • After optimization: "Explain the concept of 'tree' in computer science."

• Detailed context: If the question involves specific context or background information, provide sufficient details to help the model understand.

• Clear logic: Questions should be logically coherent with a clear hierarchical structure.

  • Before optimization: "Extract applicant name, claim amount, when the accident occurred, policy number, what documents were submitted, what kind of accident happened?"

  • After optimization: "From the claim application document, extract the following information in sequence: 1. Customer basic information: name, policy number 2. Accident-related information: accident date, accident type 3. Claim amount requested by the customer."

• Choose appropriate output method:

  • JSON output (recommended for complex, structured tasks)

    • Applicable scenario description

• Need to extract multiple fields or structured information (such as time, location, people, events, etc.)

• Need to organize extracted information for database storage or sending to other systems

• Want to verify if the output is complete or has errors (such as missing fields, incorrect types), can be used with JSON Schema for validation

• When extracted information is complex (such as an event involving multiple participants, multiple time points and locations), it is recommended to use a structured approach to organize data by categories. Below is an example showing the JSON output format defined in a Prompt for complex information extraction scenarios.

Please extract relevant information from the text and output in the following JSON format:

{
  "event_name": "Event name",
  "dates": [
    {
      "type": "Time type (such as opening ceremony, closing ceremony)",
      "date": "Date (format YYYY-MM-DD)"
    }
  ],
  "locations": [
    {
      "name": "Location name",
      "address": "Detailed address"
    }
  ],
  "participants": ["Participant 1", "Participant 2", ...]
}

• Notes

• Please make sure to include all items that need to be extracted in the JSON to ensure information completeness

• Text output

  • Applicable scenario description

    • Only need to extract a small amount of information for direct display

    • Generate summaries, notes, and other unstructured content

• Results are used for webpage display, printing, or other direct presentation purposes

• For example: Extract title, time, and location from news, generate a one-sentence summary; no need for structured processing, output in natural language form

Examples

Direct questioning checklist

Before asking, please confirm:

1. Is the information to be extracted and its characteristics clearly described?

2. Is the logic coherent, with each point clear and unambiguous?

3. Are industry jargon avoided? (Replace internal abbreviations with standard terminology)

4. Is structured output needed? If so, are all items to be extracted listed in the output format?

Advanced questioning method

Scenarios

Suitable for scenarios involving complex problems, multi-step tasks, or requiring thinking or reasoning.

Writing key points

• Concise: Even for complex tasks, express core requirements as concisely as possible, avoiding redundant explanations that may interfere with the model's understanding.

• Structured: Break down complex problems into clear steps to help the model process information in an orderly manner.

• Role positioning: Tell the model what role it should play. For example, saying "you are now a doctor" will make the model think and answer questions from a doctor's perspective.

• In-context learning (few shot): Provide the model with some good output examples for reference and learning.

• Multi-dimensional constraints: Clearly state your various requirements for the model's response, such as what the format should be and what content should be included. For example, in a financial report analysis scenario, requirements such as: amounts accurate to the nearest million dollars, use professional financial terminology, etc.

Examples

Advanced questioning checklist

Before asking, please confirm:

1. Is the core task articulated concisely?

2. Does the role setting match the nature of the task?

3. Are the thinking path and reasoning steps clearly defined? (Avoid jumping logic)

4. Is sufficient background information provided? (Complex questions require adequate context)

5. Are output format requirements clear?

6. For more complex output requirements, are clear output examples provided for the model to reference?

3. Key-Value template writing guide

Scenarios

Key-Value template is a structured data format used to clearly specify content fields to be extracted and their corresponding values. It is particularly suitable for extracting specific information from complex, information-dense long documents (such as ESG reports, company annual reports, contract documents, research papers, and technical whitepapers) and outputting in a standardized form.

Key-Value template writing techniques

Writing key points

• Precise definition: Define a clear name for each piece of information to be extracted. For example, instead of using vague terms like "basic information," use specific terms like "name" and "age."

• Semantic extension: For each piece of information (key) to be extracted, think about what other related synonyms/near-synonyms exist. For example, when retrieving "job," also consider words like "occupation," "position," etc. This expands the matching range.

• Limit output results: Through keyword options, set a clear answer range for each extraction item to constrain the model's output results.

• Clear extraction rules: Add some common questions as examples in the FAQ section to help the model understand the application scenarios and extraction methods for the keys.

Examples

Example of a contract information extraction Key-Value template

Key-value template configuration checklist

Before configuring, please confirm:

1. Is the key definition specific enough with clear boundaries? (Avoid broad and vague expressions)

2. Do the synonyms truly express the same or similar concepts? (Ensure semantic consistency)

3. Do the keyword options cover the main options for expected answers?

4. Do the FAQs simulate real user query intentions?

5. Is there any duplication or contradiction between information items (keys)? (Ensure each information item is independent)