Named entity extractor
Turning a wall of LLM prose into structured data starts with pulling out the named entities — who, where, when, and how much. This tool does browser-side named entity recognition with no model and no upload, finding persons, organizations, locations, dates, money amounts, emails, and URLs and letting you copy them as clean JSON.
A common scenario: you ask an LLM to analyze a document and it returns several paragraphs summarizing the key players, places, and financial figures. You need those values in a spreadsheet or a database table. Copying them by hand from prose is slow and error-prone. Running this extractor gives you a grouped, typed JSON array in one step.
How it works
The extractor runs two complementary passes. First, regular expressions capture the well-structured types: ISO and natural-language dates, money amounts with currency symbols, emails, and URLs. Second, a capitalization heuristic groups runs of consecutive capitalized words into candidate proper nouns, then classifies each using keyword cues — suffixes like “Inc”, “Ltd”, or “University” mark organizations, words like “City”, “River”, or known place words mark locations, and the remainder default to persons. You can filter which types are shown and export the grouped results as JSON.
What each entity type looks like in output
The copied JSON is an array of objects, each with a text and a type field, for example:
[
{ "text": "Acme Corp", "type": "organization" },
{ "text": "$4.2 million", "type": "money" },
{ "text": "March 2024", "type": "date" },
{ "text": "Berlin", "type": "location" }
]This shape is ready to feed directly into a pipeline, a database insert, or a spreadsheet import without any further parsing.
Where heuristic NER works well — and where it struggles
Heuristic extraction is fast, private, and free, but it is not a trained model. It works reliably on clean, normally-capitalized prose from standard LLM output. It struggles with:
- All-lowercase or all-caps text — the capitalization heuristic produces few person/org/location hits.
- Ambiguous abbreviations — “US” could be a country or the pronoun; context is not weighed.
- Domain jargon — industry terms like “Caspian Protocol” or “Alpha Division” look like organizations even if they are not.
For these cases, use the extractor as a first pass to catch the obvious entities and then review the JSON before using it downstream.
Tips and notes
- Structured types are the reliable ones. Dates, money, emails, and URLs are matched by pattern and rarely wrong; treat person/org/location as best-effort.
- Use it as a first pass. Heuristic NER is ideal for quickly structuring output or pre-filtering before a heavier model — not as a final ground truth.
- Sentence case matters. The name heuristic relies on capitalization, so all-lowercase or all-caps text yields fewer person/org/location hits.
- Everything stays local. No text leaves your browser, so extracting entities from private documents is safe.