Columns & AI extraction
Columns are where Quanty does its work. A column has a type, and AI columns carry a prompt: when you run the sheet, the model reads each row's documents and writes a value into every cell — with citations, confidence and a full audit trail.
Column types
When you add a column you pick its type from the Column type grid. The type cannot be changed after creation.
| Type | What it holds |
|---|---|
| File | Holds documents or files, one or more per row |
| Sub-sheet | A nested sheet inside a cell, for one to many items |
| Text | AI writes text into each cell: extractions, summaries, answers |
| Number | AI extracts a numeric value, with formatting options |
| Single choice | AI picks exactly one of the options you define |
| Multiple choice | AI picks any number of the options you define |
| URL | AI finds a link, optionally using web search |
| JSON | AI returns structured data matching your JSON schema |
Text extraction sub-types
Text columns offer focused extraction modes:
| Sub-type | Use it for |
|---|---|
| Custom Prompt | Ask anything in your own words, the most flexible type |
| Entity Extraction | Pull out names, companies, places or other entities |
| Summary | A short summary of each document |
| Classification | Sort each row into one of your categories |
| Key-Value | Extract the value for a specific field, like an invoice number |
| Date | Find a date and normalize its format |
| Number | Extract a formatted number |
| Document Reference | Reference the document's name and metadata |
Number formatting
Number columns have a Data Format section: Currency, Percentage, Date, Decimal or Integer, plus currency symbol and decimal places.
Writing prompts
The Prompt field (up to 1500 characters) tells the model what to extract. Below it, the Model picker lists available models with cost tiers (Premium, Standard, Economy), context window size and a Recommended marker.
Inputs and @-mentions
Columns can read other columns. Type @ in the prompt to mention a sibling column — for example "Compare @Contract Value with @Invoice Total". Mentioned columns appear as chips in the Inputs section, where you can also add them with + add input….
Inputs form a dependency graph: inputs run first, and editing an input column marks this column stale so you know results are out of date.
Source documents
The Source documents selector controls which file columns feed this AI column. By default a column reads the row's documents; narrow it when a sheet has several file columns with different roles.
Data source: documents, internet, or both
Every AI column has a Data source selector with three options:
- Documents (default) — the model reads only the row's files.
- Documents + internet — the model reads the row's files and runs a live web search per row, useful when a value combines internal data with current public data.
- Internet — the model reads only live web results. The column never waits for uploads; rows just need some value to key the search on.
Each row's search query comes from the prompt with @-mentions substituted by that row's values — "Current CEO of @Company" searches for each row's actual company. On text rows without mentions, the row's own content names the subject.
Web-source options: Search method (Auto — AI picks the best method per row: reading linked pages, a quick search, a thorough one, or deep research; you can also pin Fast, Thorough or Deep research), Freshness (any time up to past week) and an optional Only these sites domain filter for trusted sources like gov.pl. Sub-sheet columns can use the internet too — they research a whole dataset per row — and Quanty Connect adds premium data sources to a run. See Web research & Quanty Connect.
While a web-source cell processes, it shows Searching the web… with a globe — you can watch the data being fetched the way you watch the model think. Values from the internet list their sources in the cell details panel — title, address and publication date — and the audit trail records the exact pages and query behind every value.
Each processed row runs one web search, billed from your plan's credits on top of the AI call. The cost appears as its own Web search line on the usage page.
Knowledge Base
Attach library documents in the Knowledge Base section so the model rates against your house standards. Updating an attached document marks the column stale. See Files & Knowledge Base.
Cite sources
The Cite sources toggle adds page references to every value ("Adds page references · ~20% more credits"). Citations open in the document viewer with the source highlighted.
Branching and human review
Two settings turn a sheet into a pipeline with routing:
- Branch — run only when — the column runs only for rows where a chosen Single choice column has one of the selected values. Off-branch cells never compute (0 credits) and show not on this branch.
- Route to human review when — for Single choice columns, pick the values that should queue a row into the review panel. This acts as the end-of-pipeline gate; see Reviews & audit trail.
Sub-sheet columns
A Sub-sheet column defines Sub-columns that AI fills per item: each sub-column has a field name and a type — text, number or PY (a Python expression computed per item at 0 credits). Use it for line items, schedules, or any one-to-many structure inside a row.
AI or Formula
Every value column has a "How this column is filled" choice:
- AI — "The model reads each document and extracts the value."
- Formula — "Computes from other columns — deterministic, no credits spent." See Formulas & Excel mode.
Running the sheet
- Run AI processes all pending cells. A confirmation explains: "Pending cells will be processed with AI. This uses credits from your organization's AI budget — you can watch progress and stop at any time."
- When upstream changes made cells stale, the button offers Update, or "Run AI — also updates N out-of-date cell(s)".
- Stop halts the run; values already computed are kept.
Cell statuses during and after a run: Completed, Processing, Pending, Failed and Not on this branch. The processing bar shows ready, processing, pending and failed counts, with Retry failed files one click away.
Two states that are deliberately not errors:
- Blank rows are skipped. A row with no document, no content and no values is not work — runs leave its AI cells quietly pending instead of failing them.
- A parsing document is a wait. While a row's file is still uploading or indexing, its AI cells read Waiting for the document to be processed and run automatically once it's ready.
Every run is audited
Each cell keeps an append-only trail of every extraction run — the prompt as it was, the context used, the raw response and cost. You can restore any past value. See Reviews & audit trail.
Compare models (A/B)
From a column's context menu, choose Compare models (A/B)… to run the same prompt with two models side by side and see where they disagree before committing to one.
The pipeline map
The pipeline view renders your sheet as a graph: columns as nodes, dependencies as edges. Node kinds include AI column, Sub-sheet, Sub-sheet column, Documents, Knowledge Base, External data and Formula; edges are labeled documents, value, branch, knowledge base or external data.
Each node shows audit aggregates — Runs, Failed, Tokens, Cost, Avg. confidence and Last run — and the view offers Recompute stale, zoom controls and Export (JSON audit data, light or dark theme, optionally including prompts & formulas, audit stats and value history).
Column management
- Duplicate column copies configuration (and can copy values).
- Pin Column left or right, set Width, or Hide from view in Display Settings.
- Drag to reorder; deleting a column requires confirmation.
- Column templates save a configured column for reuse across a workbook — pick a template when adding a column instead of configuring from scratch.
Related pages
- Formulas & Excel mode — deterministic columns and grid editing.
- Reviews & audit trail — approving, refining and restoring values.