Harvey AI Vault: organize a collection and check results
Use Harvey Vault for document-based research with a clear collection boundary, testable questions and a review of the passages behind the answers.
Updated 5 min read
Harvey Vault (opens in a new tab) is a document repository for querying, reviewing and analyzing a collection of documents. To use it for research, define what belongs in the collection, ask a question you can check and inspect the source passages behind the result. fenbs can keep a generic process checklist between sessions, while client material stays in systems your firm has approved.
A useful collection has a purpose. “Find the relevant provision across this approved sample” gives a reviewer a clearer job than “tell me everything important.” Begin with a narrow task and a sample you understand before extending the method to a larger body of work.
Where do you find Vault and its setup instructions?
The product is called Vault, and Harvey’s Vault support topic (opens in a new tab) links an article titled Vault: Analyze Large Document Sets at Scale. The detailed article requires an authenticated route that was not publicly accessible during review. Use your organization’s authenticated instructions for the exact creation and upload sequence.
Access is organizational. Ask your administrator whether you have Vault access and which approved sources you may use. Do not infer that a feature is available in your deployment because it appears on the public product page. The product page describes several surfaces, but it does not establish identical capabilities or entitlements everywhere.
For choosing a tool, see Claude vs Harvey. This guide focuses on the work after that choice: preparing a research collection and judging the result. The broader AI at work hub covers the surrounding habits that make AI output easier to review.
What belongs in the research collection?
Write a short collection description in your approved working system before adding documents. State the question, the kinds of documents included and the known gaps. For a test, choose synthetic or public material the firm permits. This makes it possible to compare the answer with a known source without exposing confidential work.
- Purpose: what question should this collection help answer?
- Boundary: which documents, versions and periods are included?
- Exclusions: what material must not be added or connected?
- Ownership: who maintains the collection when a source changes?
- Review: who checks the result and what must they inspect?
Keep an inventory in the approved system. A reviewer needs to know which document is which and whether the set is complete for the question. A missing attachment or an older version can change the meaning of an otherwise accurate extraction. A larger collection is not automatically a better collection.
How should you handle permissions and synced sources?
Vault describes configurable access, viewing and editing permissions, along with sync integrations including iManage, SharePoint and Google Drive. Confirm the connection and access rules with your administrator. A source being available in another firm system does not by itself establish that it belongs in this research collection.
Before relying on a shared collection, agree who may add sources, correct the set and use its output. Test the intended access with a small approved example. Keep the test focused on the actual roles involved rather than using one highly privileged account to stand in for everyone.
If source material changes during a review, stop and establish which version supports the result. Record the remaining checks in the approved working record. Do not silently combine an earlier answer with a later document and call the combined result verified.
How do you test a question before scaling the review?
Use a small sample whose expected answers a reviewer can find manually. Include one example where the requested information is absent and another where the wording needs surrounding context. These are useful test choices, not claims about automatic test features inside Vault.
- Ask one bounded question and specify the fields you want in the answer.
- Request a source location for each extracted fact and a clear indication when information is missing.
- Compare each result with the underlying passage and its surrounding text.
- Note unsupported conclusions, omitted qualifications and ambiguous matches.
- Revise the question or collection and repeat the affected checks before expanding the work.
Vault supports review tables, but a row is still an answer to inspect. Treat a blank field as a question: was the information absent, inaccessible or simply missed? The guide to AI agents for research explains how to define a research task with a reviewable finish line.
What do you check in an extracted answer?
| Result | Review question |
|---|---|
| Extracted wording | Does it match the source and preserve any conditions? |
| Document identity | Is this the intended document and version? |
| Summary | Does it distinguish the source statement from an inference? |
| Missing information | Was the absence checked against the appropriate material? |
| Collection-wide conclusion | Does the reviewed evidence support applying it across the set? |
Repository capacity is different from what one question can process. Do not assume that every stored document was considered in a particular answer. Ask Harvey and your administrator about the relevant limits and review the evidence returned for the specific question. Avoid turning a storage claim into a promise of complete analysis.
For legal propositions beyond document extraction, use the Harvey legal research guide. For a task you want to repeat with guided inputs, see Harvey workflows. Keep those questions separate from whether this collection supports this particular answer.
Which lessons can leave the document collection?
A reusable lesson might be that a review question should ask for the exception alongside the main provision. That is a generic process improvement. A source passage, client conclusion or matter-specific reason for the improvement may be confidential. Client confidences and deal data stay out of any tool the firm has not approved.
Keep source documents and work product in approved firm systems. Harvey’s Library & History topic (opens in a new tab) describes places to access reusable prompts and previous queries; its public description does not establish an export procedure. For generic repeatable methods, the AI workflow library guide explains what makes a saved procedure useful.
Never make the same mistake twice
When something goes wrong, or right, save what you learned in fenbs. The next AI you connect reads it before it starts.
Create a free accountFree for up to 5 people and 50 open tasks · No credit card · Works with Claude, ChatGPT, Cursor and more
How to keep review lessons with fenbs
Keep only the approved general method so you can carry on from exactly where you left off, on any device and in another connected AI. fenbs is not the right place for client files, extracted matter details or confidential findings.
- Create a free account.
- Write a simple generic checklist: identify the document, check the passage, preserve exceptions and ask about missing information.
- Connect Harvey to fenbs through Integrations only if your firm approves and its administrator confirms and configures a supported connection. No ready-made Harvey connector has been verified.
- Save the safe process lesson and the next test to run. Leave out client names, source text, matter facts and confidential details.
Harvey’s Connector Library (opens in a new tab) supports administrator-controlled custom MCP connectors, but does not publish a complete setup path. The fenbs overview explains how another approved assistant can read the saved method when you return.