# bulk-annual-report-extractor

**Repository Path**: ceheps/bulk-annual-report-extractor

## Basic Information

- **Project Name**: bulk-annual-report-extractor
- **Description**: 运用大语言模型从年报文本中提取任意信息——经管类科研Skill | Extract any information from A-share annual reports using LLMs — a research skill for economics & management
- **Primary Language**: Python
- **License**: MIT
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-06-24
- **Last Updated**: 2026-06-24

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# bulk-annual-report-extractor

**Research Skill for Economics & Management — Extract Any Information from Annual Reports Using LLMs, Out of the Box**

Automatically downloads annual reports of A-share listed companies and extracts structured information from the report text using large language models (DeepSeek), with verifiable source evidence. Produces panel data ready for empirical research, eliminating tedious manual data collection.

## Who Is This For

Researchers and students in economics and management — no programming skills required.

## Quick Start

### Prerequisites

Install any mainstream Agent product: [WorkBuddy](https://workbuddy.ai), [Codex](https://github.com/openai/codex), [Claude Code](https://claude.ai), etc.

### Install Skill & Environment

Open your Agent and send the following instruction:

> Help me download the bulk-annual-report-extractor skill from https://github.com/Ceheps/bulk-annual-report-extractor and set up the required environment.

Then follow the Agent's prompts:
- Allow the Agent to install Python dependencies
- Enter your [DeepSeek API key](https://platform.deepseek.com/api_keys) ([how to obtain](https://platform.deepseek.com/api_keys))

### Extract Information

Tell your Agent:

> Use the bulk-annual-report-extractor skill to download company annual reports based on my requirements and complete structured information extraction.

Under the Agent's guidance:
1. Specify target companies and year range
2. Define the data fields to extract (e.g., innovation goals, information redundancy, etc.)
3. Review the sample results provided by the Agent, then confirm to proceed with full-scale processing

**Cost Note**: The skill uses DeepSeek V4-flash by default. Measured cost is approximately **¥0.02 per report**.

## Workflow

```
Download Reports → PDF to Markdown → Chapter Structuring → Smart Chapter Targeting → LLM Extraction → Traceable Excel Export
```

## Output Files

| File | Description |
|------|-------------|
| `result.xlsx` | Full extraction results with source traces (chapter, original text excerpt) |
| `result_clean.xlsx` | Values-only panel data, ready for empirical analysis |
| `validation_sample.xlsx` | Validation sample for manual review |

## Technical Features

- Batch processing of **3,000+** annual reports
- Structured segmentation + targeted extraction (not full-text LLM feeding), significantly reducing costs
- Traceable results (each extraction comes with source evidence)
- Resumable — continue from breakpoints after interruption
- Hallucination filtering — automatically discards extractions that cannot be grounded in the source text