Welcome to OmniExtract

OmniExtract is an LLM-based automatic extraction tool based on DSPy, specifically designed for information extraction tasks from literature and documents. It utilizes prompt optimization engineering to enhance extraction performance based on curated data, and provides various file format parsing tools, supporting batch extraction of multi-property entities from original documents (in formats such as PDF or XML) and tabular files. A video tutorial about OmniExtract is available at OmniExtract - Tutorial.

Online Version

This is the online version of OmniExtract for test. It provides simplified document information extraction, prompt optimization, and table information extraction functions.

Current Model Configuration:

Document Information Extraction & Prompt Optimization: Qwen2.5-72B-Instruct
Table Information Extraction: Qwen3-235B-A22B-Thinking-2507

Note: These models will be replaced with more modern ones in the near future, once information extraction testing is complete.

Daily Usage Limit:

Due to computational resource constraints, we have set a daily limit on the number of processing tasks.
Remaining tasks today: 50 / 50
Limit resets in 22 hours 43 minutes

We recommend visiting https://github.com/wyb39/OmniExtract to start the local service, where you can use more customized functions.

Recommended Prompt Template

We recommend using the following template to clearly describe the criteria that the information extraction task should follow.

Extract [target information] from [source].
Follow these rules:
1. [rule 1]
2. [rule 2]
...
Exclude the following:
1. [exclusion 1]
2. [exclusion 2]
...

Example

Extract clinical information for patient and healthy control groups from biomedical literature.
Follow these rules:
1. Preserve full group names exactly as reported.
2. When extracting age, report values in years (convert units to years when necessary).
3. When calculating sex percentages, report values rounded to two decimal places.
4. Keep all extracted values precise and cite the source sentence or table.
Exclude the following:
1. Non-human studies.
2. Review articles and meta-analyses.

Example Packages

Each package bundles the config file, the test data, and a README for a complete example. Download a package, unzip it, and follow its README to try the corresponding module.

Document Extraction

Original prompt: Runs Document Extraction - Original on AKC breed standard PDFs. Includes the config, the test PDFs, and the README. Download (.zip)

Optimized prompt: Runs Document Extraction Optimized on AKC breed standard PDFs. Includes the config, the optimization package, the test PDFs, and the README. Download (.zip)

Prompt Optimization

Breed prompt optimization: Optimizes the extraction prompt from curated AKC dog-breed data. Includes the config, the curated dataset, the PDF files, and the README. Download (.zip)

Table Extraction

EWAS table curation: Extracts methylation probe data from EWAS supplemental tables. Includes the config, the test tables, and the README. Download (.zip)

Datasets used for performance evaluation

critical_cooling_rate_truth.xlsx: Ground-truth data used for evaluating critical cooling rate extraction quality. Download (.xlsx)

ewas_related_publication_list.xlsx: Reference publication list for EWAS-related evaluation tasks. Download (.xlsx)

ewas_table_extract_result.xlsx: Extraction result benchmark table for performance comparison. Download (.xlsx)

wiki_recycled_dataset.zip: Packaged dataset collected for wiki-related recycling evaluation. Download (.zip)

yidu-s4k_dataset.zip: Dataset bundle used for model performance evaluation. Download (.zip)