Every quote, checked against the text. 330 of 332 facts on a government contract carry the exact words they were read from.

short-form-contract.md

Loading…

Extractor
short-form-contract.md

Loading…

Extractor
Request a demo

Swipe up to continue

Starts the moment pages arrive. Each window of text goes to the model as soon as its pages are in: on a spreadsheet, 45 windows 13 ms after the sheets arrived, side by side.

cpws2022-extract.xlsx · timeline
Pages in: 0.33 s, 0.33 sFirst window sent: 0.35 sWindows: 45
0 s34 s68 s
Request a demo

Swipe up to continue

Quotes are verified, word for word. Each relationship carries a word-for-word quote. A quote that is not in the text is dropped.

Block p1e7

SwiGLU activation function [PaLM]. We replace the ReLU non-linearity by the SwiGLU activation function, introduced by Shazeer (2020) to improve the performance. We use a dimension of \(\frac{2}{3}d\) instead of 4d as in PaLM.

relationship.json
{
  "type": "REPLACES",
  "from": "SwiGLU activation function",
  "to": "ReLU non-linearity",
  "evidence": [
    "We replace the ReLU non-linearity by the SwiGLU activation function"
  ],
  "block": "p1e7",
  "page": 1,
  "quote_found": true
}
Request a demo

Swipe up to continue

A second pass for what was missed. Names, numbers, dates and identifiers the first pass skipped go back to the model by name.

extractor · checks

Checklist: 16 of 85 item(s) not accounted for after the first pass; second pass on 2 chunk(s): 4 extracted, 11 set aside as carrying no information, 1 unresolved (such as: 2,000)

Extraction coverage: 47% of the text is quoted as evidence; 120 of 120 fact(s) carry a verified quote; 2 of 15 element(s) below 30% (least covered: p1e9 0%, p1e14 24%)

Request a demo

Swipe up to continue

No schema to write first. The model names the labels and relationships, then labels that mean the same are merged across the document. Your experts review the draft in Sculpting Studio.

labels_merged.json
{
  "llama-page3.pdf": {
    "Architecture": [
      "Model"
    ],
    "ModelFamily": [
      "Model"
    ],
    "ModelSize": [
      "Model"
    ],
    "Optimizer": [
      "Technique"
    ]
  },
  "cpws2022-extract.xlsx": {
    "Location": [
      "Region"
    ],
    "SurveyResponseCategory": [
      "AnswerOption"
    ]
  }
}
Request a demo

Swipe up to continue

100,000-row spreadsheets, mapped by column. A large table is described by its columns and mapped once, then the mapping runs over every row in code.

dispensing-2025.csv · 100,000 rows
monthpharmacy_idpharmacyregiondrugitems
2025-01FA102Northgate PharmacyNorth EastAmoxicillin 500mg412
2025-01FA102Northgate PharmacyNorth EastMetformin 500mg1,208
2025-01FC318Riverside ChemistLondonAtorvastatin 20mg966
2025-01FC318Riverside ChemistLondonAmoxicillin 500mg287
2025-02FK044Market Square PharmacySouth WestMetformin 500mg734
2025-02FK044Market Square PharmacySouth WestSalbutamol inhaler158
+ 99,994 more rows
column-mapping.json
{
  "rows": 100000,
  "entities": {
    "Pharmacy": {
      "key": "pharmacy_id",
      "name": "pharmacy"
    },
    "Region": {
      "name": "region"
    },
    "Medicine": {
      "name": "drug"
    }
  },
  "relationships": [
    {
      "type": "LOCATED_IN",
      "from": "Pharmacy",
      "to": "Region"
    },
    {
      "type": "DISPENSED",
      "from": "Pharmacy",
      "to": "Medicine",
      "properties": {
        "items": "items",
        "date": "month",
        "date_precision": "month"
      }
    }
  ],
  "model_calls": 1,
  "applied_in_code": "100,000 rows in 1.2 s"
}
Request a demo

Swipe up to continue

A graph you can query. Every relationship comes with its dates, its quote and the block, page and section it was read from.

relationship.json
{
  "type": "MUST_NOTIFY",
  "from": "Supplier",
  "to": "Buyer",
  "valid_within": "3 Working Days",
  "date_text": "within 3 Working Days",
  "evidence": "the Supplier must notify the Buyer within 3 Working Days of the Supplier becoming aware of the Material Breach",
  "block": "p1e81",
  "page": 1,
  "section": "7 Record keeping and reporting"
}
Request a demo

Swipe up to continue

More about Extractor. Speed, verified quotes and what it takes to run it.

Request a demo

Swipe up to continue

Benchmarks

Detailed, and fast where it counts. The time goes to the models; everything else takes a fraction of a second.

Facts with a verified quote

330 / 332

On a government contract; 754 of 841 on a wide spreadsheet.

20-page document

<15 s

Once read, end to end.

Graph write, 300 pages

0.39 s

8,403 nodes and 24,599 edges.

100,000-row sheet

1.2 s

Read and laid out.

Built for production

Built for production

An outline from the document itself

The real table of contents when there is one, checked against the pages.

outline.json
[
  {
    "structure": "1",
    "title": "2.2 Architecture",
    "level": 1,
    "pages": [
      1,
      1
    ]
  },
  {
    "structure": "1.1",
    "title": "2.3 Optimizer",
    "level": 2,
    "pages": [
      1,
      1
    ]
  },
  {
    "structure": "1.2",
    "title": "2.4 Efficient implementation",
    "level": 2,
    "pages": [
      1,
      1
    ]
  }
]

Dates you can sort

Normalised with their precision, the wording kept.

relationship.json
{
  "type": "COVERS",
  "from": "community pharmacies",
  "to": "England",
  "count": "11279",
  "date": "2021",
  "date_text": "At the time of the 2021 survey",
  "date_precision": "year",
  "evidence": [
    "there were 11,279 community pharmacies operating in England"
  ]
}

One thing, a group or a kind

A kind of thing is never merged with one instance of it.

short-form-contract.md · entities
One thing105 in the contract
Short Form Contract for the Supply of Goods and/or ServicesDocumentSpecificationDocumentAnnex 4 – Supplier Tender (Optional)DocumentClause 4.1Clause
A group8 in the contract
BuyersPartyRoleAll GoodsDocumentSubcontractorsPersonSupplier StaffPerson
A kind24 in the contract
SupplierPartyRoleBuyerPartyRoleDeliverablesDocumentGood Industry PracticeConcept