Extraction

Without extraction, an uploaded file stays a file. The project graph does not receive entities or triples from it.

One file often mixes two shapes of information: repeating records (a parts table, a CSV) and free-form prose (a procedure, a cover letter). If yggy treated every page as one blob of text, tables would lose their columns. If it treated every page as a table, the narrative would be ignored.

Extraction reads the file, chooses a path from that structure, and writes facts into the active project’s graph.

What you get when it finishes

Open Sources. When the status is Completed, the document can show:

• graph facts on Triples;
• a Mapping tab when repeating fields were used;
• pending identity when identifier fields still need approval;
• Completed with issues when part of the file produced no graph data.

You do not pick Structured, Text, or Mixed. yggy chooses from the file. Overview shows the method that actually ran.

How yggy chooses a path

flowchart TD
  Upload[Upload a file] --> Read[Read text and images]
  Read --> Detect{Tables or repeating records?}
  Detect -->|Yes| Structured[Structured extraction]
  Detect -->|Also prose left| Text[Text extraction]
  Detect -->|Mostly prose| Text
  Structured --> Combine[Combine facts]
  Text --> Combine
  Combine --> Label{What landed?}
  Label -->|Repeating fields only| StructuredLabel[Structured SGSE]
  Label -->|Prose only| TextLabel[Text extraction]
  Label -->|Both| MixedLabel[Mixed]
  Combine --> Graph[Write the project graph]

1. Read the file. yggy extracts text. Scanned pages and images can use OCR.
2. Detect structure. Tables and repeating records can follow the structured path. Surrounding prose can follow the text path.
3. Extract facts. Structured extraction designs a field mapping and applies it to every row. Text extraction reads the remaining prose.
4. Resolve entities. yggy compares proposed entities with existing project data.
5. Write the graph. Accepted entities and triples link back to the document.

Labels on Overview

Label on Overview	What it means	Typical source
Structured (SGSE)	Repeating fields became graph facts through a saved mapping.	CSV, spreadsheet, or a detected table.
Text extraction	Free-form prose became graph facts.	A contract PDF or a notes file.
Mixed	Both paths produced facts in the same document.	A service manual with a parts table and a procedure chapter.

Manual is a fact someone added in the product, not an automatic extract.

Structured extraction (SGSE) means: the model sees a small preview and writes a mapping. Code then applies that mapping to every record. The mapping is what you edit on the Mapping tab. See Mappings.

Text extraction means: yggy reads the prose in sections and proposes triples from the sentences.

Mixed means both happened. Each triple still shows its own source on Triples. The document label is Mixed when both kinds of facts remain after they are combined. Structured facts win if both paths extract the same statement.

Example

Upload pump-service-manual.pdf. Page 2 is a parts table (part_no, name, fits_model). Pages 3–8 are torque and inspection steps.

yggy maps the table through structured extraction and reads the steps through text extraction. Overview shows Mixed. Mapping shows how part_no becomes the part entity. Triples lists both part_of facts from the table and requires facts from the procedure.

A file that is only suppliers.csv shows Structured (SGSE) and has a Mapping tab. A file that is only cover-letter.docx shows Text extraction and has no Mapping tab.

After extraction

• Use Re-extract triples after you edit a mapping or when you want a new run. Re-extraction replaces the current triples for that document. When the vocabulary and mapping still match, yggy can reuse the mapping instead of designing a new one.
• A completed extract can still need a person. Unapproved identifier fields show pending identity on the document and on Ontology → Identities. New vocabulary can wait in Ontology review while the rest of the graph stays usable.
• Extraction is probabilistic. It can miss facts, split one entity, or infer the wrong relationship. Each fact on Triples can show a source quote. Content shows the extracted text and tables so you can read what yggy saw. Neither makes every fact correct.

Track status and actions in Extraction status. Graph Intelligence throughput follows the organization plan. See Billing.

Inference in organization settings chooses where models run (cloud or Mímir). It does not switch Structured, Text, or Mixed. That choice still comes from the file.