Extract named entities from a PDF — people, organizations, places, dates as structured data
Named entity extraction surfaces the proper nouns and date references in a document as structured tables. People, organizations, places, dates — each in its own table with page citations. Useful for any workflow that needs to know "who and what is mentioned" in a long document. Here is what the extractor surfaces, the four workflows where the structured view replaces manual scanning, and the precise difference between this AI tool and the free regex-based Extract Contacts.
What the extractor produces
Drop a PDF. The tool reads the text, runs an AI named-entity-recognition pass, and produces four Markdown tables — one each for People, Organizations, Places, and Dates. Each row has the entity name, a one-line role note ("CEO of Acme Corp," "city in Karnataka," "founding date"), and the source page where the entity appeared. Multiple mentions of the same entity collapse to a single row with multiple page citations.
The output is structured for downstream use: copy a table directly into a spreadsheet, import to a CRM, paste into a research-tracking system, feed into a knowledge graph builder. The page citations make every entity verifiable against the source.
Four workflows where structured entity tables earn their place
Cases where you need entities as data, not as in-flow text:
- Research paper preprocessing. Building a literature review or citation graph. The People table gives you authors mentioned in citations; the Organizations table gives you research institutions; the Places table gives you study locations. All ready for downstream analysis.
- Legal-document discovery. Court orders, contracts, and litigation documents reference many people and organizations. Extract Entities surfaces them all in one pass for case-management ingestion. Use the page citations to verify every entity against the source.
- Business-intelligence on company reports. Annual reports mention partners, suppliers, customers, competitors. Extract Entities surfaces the relationship landscape from the report's narrative.
- Compliance/KYC document review. When reviewing a long document for KYC compliance, the People and Organizations tables surface every named party. Combined with a sanctions-list cross-check, this is the operational starting point for due diligence.
AI Entity Extraction vs free Extract Contacts
Two adjacent tools that solve different problems:
- Extract Contacts (free, regex). Finds emails, phone numbers, URLs. Pattern-based — explicit string formats. Misses obfuscation. Free.
- Extract Entities (AI, 3 credits). Finds named people, organizations, places, dates. Context-based — entities recognized by what they ARE, not by what they look like. Catches "Mr. Sundar Pichai" as a person and "Alphabet Inc." as an organization without needing either to follow a regex pattern. Costs credits.
How accurate is the extraction
Three factors affect accuracy:
- Common Western names: very high. Names like "Sundar Pichai," "Microsoft," "Bangalore," "2024" are reliably identified across document types.
- Less-common Indic names: high but variable. Names with non-standard transliterations or regional variants sometimes mis-classify. The Sundar Pichai example works because the name is well-known; a regional-newspaper byline with a rare regional name may classify less reliably.
- Ambiguous proper nouns: medium. Words that are both common nouns and proper nouns ("Apple" the fruit vs "Apple" the company; "Bishop" the chess piece vs "Bishop" the surname) get disambiguated by context, but errors happen. Spot-check ambiguous entries.
Relationships and what the tool doesn't do
Three explicit non-features:
- Doesn't infer relationships. Two people mentioned in the same document aren't claimed to know each other unless the source explicitly says so. For relationship inference, build a knowledge graph downstream from the entities + the source text.
- Doesn't classify entity sub-types. Person table doesn't separate "executives" from "academics" from "defendants." Organization table doesn't separate "customers" from "vendors" from "competitors." The role-note column captures some of this in narrative form but isn't structured.
- Doesn't resolve to canonical IDs. "Sundar Pichai" doesn't get a Wikidata ID; "Microsoft" doesn't get a CIK or stock ticker. For canonical resolution, downstream tooling against a reference database is required.
Three patterns for working with the output
Habits that improve downstream use:
- Dedupe before downstream use. The tool collapses exact-string duplicates within a document but doesn't normalize variants. "Sundar Pichai" and "Mr. Pichai" may surface as separate rows. Quick post-processing dedupe is usually needed.
- Sanity-check ambiguous entries. Click through to the citation page for any entity that surprises you. AI is reliable but not infallible; spot-checks catch the rare misclassification.
- Combine tables for graph construction. Person + Organization + Place tables together describe the document's named-entity landscape. Build a knowledge graph by treating co-mention in the same page as an implicit edge between entities.
Limits and pricing
Extract Entities charges 3 credits per document. The tool handles PDFs up to 25 MB. Processing runs on our servers; the document is in memory only during extraction and is never persisted. Output is four Markdown tables with page citations.
Common pairings: Extract Entities + Extract Contacts for the full named-entity + contact-method picture. Extract Entities → Knowledge graph downstream. Extract Entities + Sentiment Analysis to see which entities are associated with positive vs negative sentiment in the source.