Skip to main content
For AI agents: a documentation index is available at https://docs.coverbase.com/llms.txt. This page is also available in markdown by appending .md to the URL.
This guide is part of the User Guides collection. It runs through the Coverbase MCP server, so you need a connected assistant. Setup is in Connecting.
A document evidence pull answers a question out of a vendor’s stored documents. You ask something like “does the SOC 2 cover encryption of data at rest?” and Coverbase decides which one or two documents are worth reading, reads them, and comes back with the passage that answers you. The answer comes with its source attached: the sentence as it appears in the document, the page it sits on, an image of that page with the sentence marked, and a link that opens the document there.
This is different from Document Insights. Insights are fields you configure once and Coverbase populates on every matching document. An evidence pull is a one-off question you ask in the moment, about anything, with no configuration.

​
What you get back

A direct answer

One or two sentences answering the question, per document read.

A verbatim quote

The exact span from the document that supports the answer, checked against the document text before it is returned. If no quote could be grounded, the answer says so and you should treat it as unverified.

A page number

Where the quote sits in the document. Excel workbooks report a sheet name instead.

A snippet image

A crop of the page with the quote highlighted, rendered during the pull and displayed inline in chat.

A highlighted deep link

Opens the document in Coverbase on the right page with the passage marked. Details in The highlighted link.

A confidence

High when the document states it outright, medium when it had to be inferred, low when the model was unsure.

​
Running one

​
The scripted way

Invoke the pull_document_evidence prompt from your client’s prompt menu and give it a vendor name and a question. It runs the whole sequence: resolve the vendor, look at what documents exist, start the pull, wait for it, and report each answer with its quote, page, image and link. See Canned prompts.

​
The conversational way

Ask in plain language. All of these start a pull:

A question about one vendor's evidence

Does Acme’s SOC 2 cover encryption of data at rest?
The assistant resolves Acme, sees which documents you hold, and starts a pull scoped to that vendor. If several vendors match the name it asks which one rather than guessing.
What is the notice period for termination in Acme’s MSA?
Naming the kind of document narrows the candidate set before anything is read. You do not have to: the selection step is usually right without it. It helps most when a vendor has dozens of documents and the question points at one type.
In the Acme penetration test report, were any critical findings left unresolved?
When you name a document, that document is what gets read. No selection step.
For the Acme Data Warehouse service, does any document say where data is stored?
Restricts the candidate set to documents attached to that service.
What did Acme’s most recent SOC 2 say about change management? Only look at documents uploaded this year.
The pull itself has no upload-date filter. The assistant uses upload dates when it lists the vendor’s documents, and then points the pull at the document it found, or you can name that document yourself.

​
It takes a couple of minutes

Starting a pull returns a run id, not the answer. The file is fetched, a model reads it, and the passage is located and rendered. A typical pull finishes in one to three minutes, and a long document can take longer. Your assistant should tell you it is running and then collect the result. If it goes quiet, ask it to check the run.
Do not start a second pull because the first has not come back. Each pull reads documents at full cost. Ask for the result of the run you already have.

​
A worked example

You ask:
Does Acme encrypt customer data at rest? Check their SOC 2.
The assistant starts the pull and tells you so:
A minute or two later it collects the result:
The ![Evidence snippet] line renders in chat as a crop of the page with the passage marked:
A crop of a SOC 2 report page. Under the heading CC6.7 Encryption, the sentence beginning All customer data is encrypted at rest using AES-256 is highlighted in yellow across two lines. The unhighlighted CC6.1 Logical Access and CC7.2 Monitoring controls sit above and below it.

The snippet image returned with the answer: the quoted passage highlighted in place, with the surrounding controls left visible for context.

Cite the quote and the link rather than the answer. They are what show where the answer came from.

​
How Coverbase decides what to read

Reading every document a vendor has for every question would be slow and expensive, so a pull narrows in stages.
1

Filter on metadata

Your document type and service filters, or the one document you named, are applied in the database first. Nothing is read yet. At most 100 documents survive as candidates.
2

Gather what was already quoted

Coverbase collects passages that earlier work already quoted from those same documents, along with the images already rendered for them. This is what makes a repeat question cheap.
3

Let a model pick

A model sees the candidate list (names, types, upload dates) and those earlier passages, and picks the one or two documents most likely to hold the answer. It also decides whether an earlier passage already answers the question outright. Its reasoning comes back to you as the Selection line.
4

Read only those

The chosen documents are read in full. By default at most two, and never more than five.
5

Locate and render the passage

Each quote is located in the page and cropped with the passage highlighted. If a matching image was already rendered for that passage, it is reused rather than made again.
Read the Selection line when an answer is not what you expected. Usually a different document was picked than the one you had in mind, and the line says which and why. Name the document explicitly to override it.

​
Asking the same thing twice is cheap

When a passage has already been quoted and rendered, a later pull can answer from it without reading the document again. You will see this in the header:
Read 0 means nothing had to be opened. The finding will say Source: a passage already cited from this document, and the snippet image is the one rendered earlier. The answer is grounded in the same verbatim text either way. Coverbase never reuses a passage whose underlying insight has since been re-extracted, so a stale image cannot ground a new answer.
Every answer carries a link of this shape:
Opening it loads the document preview, jumps to that page, and marks the passage. The highlight travels as the quote itself rather than as coordinates on the page. If the document is re-processed and the layout moves, the viewer re-matches the text and still finds it, where a stored box would point at the wrong place. It also keeps the link readable, so you can see what it will highlight before you click.
The link opens Coverbase and respects your permissions. Someone without access to that vendor sees your normal access error, not the document. Sharing the link with a colleague is safe.
The snippet image in the chat response is a short-lived link that expires about an hour after the result is collected. The Coverbase link does not expire. When you paste evidence into a ticket or a report, paste the Coverbase link, and save the image itself if you need it to persist.

​
What can and cannot be read

Scanned documents work

Coverbase OCRs documents on ingestion and keeps a searchable rendition alongside the original. A pull reads that rendition, so a scanned PDF that has finished processing is readable and its passages can be highlighted. Quality follows the scan: a clean scan reads like a native PDF, a photographed page at an angle may not.
A workbook is read as a workbook rather than as its PDF rendition, because converting to PDF loses the cell structure. Answers from a workbook cite a sheet name instead of a page number, and do not carry a snippet image.
A document that has not finished ingestion has no readable rendition yet, so it is not a candidate. Wait for it to finish and ask again.
A document past the live-analysis size limit reports Could not read this document with the reason, rather than silently dropping out of the answer.
If a document was read and does not address the question, the finding says No answer to this question in this document. This usually means the clause you expected is not in the vendor’s paper.

​
Reading the answer critically

An answer without a quote is unverified

When the model could not ground its answer in a verbatim span, the response says: No verbatim quote was grounded in the document, treat the answer as unverified. Open the document and check it yourself before using the answer.
The header says how many documents were read out of how many were in scope. Read 2 of 40 means thirty-eight documents were not opened. The answer is what those two say, not a statement about your whole repository.
Medium or low means the model had to infer. The quote is still verbatim; what is uncertain is whether that quote means what the answer says it means.
“None of the documents read answered the question” is not “the vendor has no such control”. Drop the document-type filter, raise the document limit, or ask about a specific document you believe holds the answer.

​
Limits

A pull that has not reported back within about 25 minutes is reported as stalled. The first pull will not return a result, so start another one.

​
Permissions and audit

A pull reads only documents you can already see. It runs under your organization and your permissions, exactly like opening the document in the dashboard. Asking a question does not require write access. The pull records its own question and answer so the result can be collected later, and that record is the only thing it writes. It creates nothing you manage elsewhere: no finding, no note, no document. Every pull is captured in the MCP audit log with the tool called, the user, the organization, and the timestamp. See Permissions and Security.

Document Insights guide

Fields pulled from every document of a type automatically, with the same quote and page citation.

Evidence quality

What makes a document usable as evidence in the first place.

Canned prompts

pull_document_evidence and the nineteen other scripted workflows.

Tool reference

The exact routes behind a pull, for auditing what an assistant did.