When the Answer Depends on Every Document

Search finds the few passages closest to a question. Many of the questions a firm asks need every document read: how many, which ones, and what each one says.

The Questions Search Gets Wrong

A partner pricing a new engagement asks what the firm proposed on every similar pitch. A COO asks which client contracts still carry the old limitation of liability. A diligence lead asks which agreements in the data room have a change of control clause.

Search, AI search included, answers by finding the passages that look most like the question and writing from them. That works for "what did we say about pricing in that healthcare proposal". It goes wrong, without saying so, on questions about the whole set. When the answer depends on every document, a few passages are a sample, and the answer reads just as confident whether it covered the whole archive or a corner of it.

So the question goes to an analyst. Open each file, copy the clause or the fee into a spreadsheet, add it up. It takes days, it is out of date soon after, and nobody wants to do it twice.

How Verra Handles It

Attach a knowledge base or a folder to a chat and ask your question. Verra works out what kind of question it is and takes the route that fits.

• Exact matches. When the question turns on exact words, a clause number, a defined term or a client name, Verra matches them across every attached document and counts documents, not passages. A contract that mentions the term many times counts once.
• Judgment calls. When the answer takes judgment, like whether a contract caps liability however the clause is worded, Verra reads each document and writes the answer into a table, with a row for each document and a column for each question. These are the same tables as Tabular Extraction in the knowledge base, so you can build one there or ask for one in a chat.
• Totals and comparisons. Once the answers are in the table, Verra adds them up, ranks them and compares them by any column, or folder by folder.

The routes chain. Matching narrows the candidates, reading settles each one, and the calculation gives the answer.

Answers You Can Check

A firm-wide number needs its documents behind it.

• Every answer names the document it came from. Open a row and the document shows beside its answers.
• Verra says which documents a number covers, and which it couldn't read and why.
• When a document doesn't state a value, the table says so. It is never counted as zero.
• Answers Verra is less sure of, or whose evidence disagrees, are flagged for review. For columns with set answers, you can see how confident it is in each.
• Ask Verra to check any answer, and it goes back to the document and quotes the passage with its page.
• Export the table to Excel: one row per document, a link to open each one, and numbers and dates typed as numbers and dates.

Reusing the Work

Tables stay. The next chat that needs the same answers reuses them instead of reading the archive again, and teammates with access to the documents find them in the knowledge base. When documents are added or changed, running the table again reads only the new and changed ones.

Verra works only on the knowledge bases and folders attached to the chat. A number about one folder is reported as one folder, never as the whole knowledge base.

What It Changes

Some questions that used to take an analyst days now get answered in a chat.

• Proposals. Every past fee on similar work, side by side, before the price goes out.
• Diligence. Every agreement in the data room checked for the same clauses, with the ones that couldn't be read listed instead of skipped.
• Knowledge management. The archive tagged by service line, client, industry and engagement type in one pass, as one table that teammates with access can reuse.
• Risk. Which engagement letters are missing a clause, by client and by year.

Search is still the right tool for finding a passage. For questions about the whole archive, Verra opens each document and keeps the answers in a table.