> ## Documentation Index
> Fetch the complete documentation index at: https://docs.coverbase.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Document Insights
> Org-defined fields extracted from every document, with a verbatim source quote, page citation, and confidence on each value.
For AI agents: a documentation index is available at [https://docs.coverbase.com/llms.txt](https://docs.coverbase.com/llms.txt). This page is also available in markdown by appending .md to the URL.
Document Insights turns any document your organization collects into structured, queryable data. You define the fields once (an effective date, a liability cap, a breach-notification window, an audit period), and Coverbase populates them automatically whenever a matching document is processed. Every value carries a verbatim source quote, a page citation, and a confidence level, so a reviewer can verify it without opening the file.
Insights are the field-level layer of contract intelligence. [Contract Guardian](/products/contract-guardian) is the language-level layer, comparing vendor clause language against your playbook. They run over the same documents and feed the same records.
## What it does
You decide what gets extracted. Each insight has a name, a description that instructs the model, a value type (text, number, date, yes/no), and a mode. No fixed schema to work around.
Every value is verified against the document before it's saved. The supporting quote must appear in the source as a contiguous span, and the cited page is corrected to where the quote actually lives.
Insights apply to the document types you choose, individually or through groups that share scope. A liability-cap field runs on MSAs, not on SOC 2 reports.
**Extract** copies a value verbatim from the page. **Synthesize** answers an interpretive question and cites the language that supports the answer.
## Extract vs Synthesize
The mode determines who composes the value: the document, or the model.
| Mode | Behavior | Use it for |
| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------- |
| **Extract** | The value is copied verbatim from the document. Coverbase verifies that both the value and its supporting quote appear in the source. If they don't, the insight returns empty rather than guessing. | Dates, party names, jurisdictions, printed figures, named frameworks: anything you could highlight with a marker |
| **Synthesize** | The model composes an interpreted answer and must cite a verbatim quote supporting it. Returns empty when the document can't answer. | Yes/no judgments, rollups across sections, normalized verdicts, values the contract implies but never prints |
A **yes/no** insight in Extract mode only succeeds if the document literally contains "Yes", "No", "True", or "False". Contracts don't. Set boolean insights to Synthesize.
The [Document Insights guide](/user-guides/document-insights) covers mode selection in depth, along with a starter library of insight definitions to copy.
## What each value carries
| Field | Description |
| ------------ | --------------------------------------------------------------------------------------------------------------- |
| Value | Stored in its declared type. Dates normalize to ISO `YYYY-MM-DD`; numbers strip currency symbols and separators |
| Source quote | The exact span from the document supporting the value, verified to exist in the source |
| Page number | 1-indexed physical page containing the quote |
| Confidence | `high` when explicitly stated, `medium` when inferred, `low` when uncertain |
| Reasoning | Notes on ambiguities or assumptions made during extraction |
| Source type | `ai_extracted` or `manual`, so hand-corrected values are always distinguishable |
## How it runs
When a document finishes processing, every enabled insight matching its type is extracted in the same pass. Insights grouped by mode run as separate calls.
Re-extract a whole document or a single insight after you add or reword a definition.
A new definition starts empty on documents already processed. A backfill re-extracts them in bulk, scoped by insight, document type, or vendor, with a cost estimate before it runs.
PDFs are read natively, including their layout. Long documents are chunked with page overlap and results merged, preserving original page numbering. Excel workbooks are grounded per worksheet.
Insights roll up onto the contract record, feed the contract AI summary, and export to CSV with every value, citation, and confidence intact.
## Auditing coverage
Defining insights is the easy half. Knowing whether they *worked* is the half that decides if anyone trusts the data.
**Configuration → Document Insights** is the module for that. It shows the same set of extractions through three lenses:
| Lens | Answers |
| ------------------ | ------------------------------------------------------------------------------------------------------------- |
| **Insights** | What have we defined, and is each definition healthy? |
| **Document types** | Is this document type well covered? A matrix of document type against insight, with a fill rate in every cell |
| **Extractions** | Is this specific value right? Every extraction in the org, filterable by confidence, groundedness, and source |
### The number that matters
A blank cell in the coverage matrix means one of two opposite things, and the module keeps them apart:
Extraction ran and found nothing. Either the definition is ambiguous, or the term genuinely isn't in that document. This is a **prompt problem**. Open the Extractions tab and read what it actually returned.
The insight was created after the document was last processed, so it has never run against it. This is a **backlog problem**; one backfill fixes it. The matrix marks these as *stale*, and the badge is the button.
## Promoting insights into custom fields
An extraction lives on a document. A custom field lives on the record people actually work from: the vendor, the contract, the assessment. Promotion connects them.
Link an insight to a custom field and Coverbase fills that field from the extraction whenever a matching document is processed, then backfills the field across every document you already hold. The field keeps a pointer to the document, page, and verbatim quote it came from, so a value on a contract record can show you *why* it holds that value.
| Insight type | Fills a custom field of type |
| ------------ | ---------------------------- |
| Text | Text |
| Number | Number |
| Date | Date |
| Yes/No | Yes/No |
Select and multi-select fields are not promotable yet. Mapping free text onto a fixed option list needs the options fed into the prompt, not matched after the fact.
Two gates decide what gets promoted, and one rule is absolute:
* **Require verified evidence** (on by default) skips any answer whose supporting quote could not be located in the source.
* **Minimum confidence** skips answers below the bar you set.
* **A value a person typed is never overwritten.** Not by a backfill, not by a later upload, not by any conflict policy. Automated writes only ever replace other automated writes.
## Common workflows
Define the fifteen fields your team asks of every agreement (value, effective date, term, renewal, notice period, liability cap, governing law, data return, audit rights) and get them populated on every contract as it lands, instead of maintaining a spreadsheet by hand.
Extract audit periods, report scopes, certification expiry dates, and auditor names from SOC 2s and ISO certificates so evidence freshness can be tracked without a human reading each report.
Once a field exists on every contract, portfolio-wide questions become answerable: which agreements auto-renew inside 60 days, which permit customer data in model training, which cap vendor liability below one year's fees.
Any value is editable. A corrected value is marked as manual with the user recorded, so a reviewer can see at a glance how much a document needed hand-checking.
Insight values surface on the document, on the contract record, in the contract AI summary, on any custom field you promote them to, and through the [MCP server](/mcp/overview) for natural-language querying.