This article is for Everlaw users who are searching their Everlaw documents in Claude via our MCP connection. The article covers some best practices and tips for formatting your searches within Claude.
To learn more about the MCP connection, and how to set it up, see Setup Steps to UseClaude to search your Everlaw projects.
Overview
Once you have connected Claude to Everlaw, you can start asking questions and Claude will search your Everlaw data to answer them. Because Claude can search, analyze, and interpret your documents' contents and metadata, there is a lot of flexibility in how you use it. You can ask:
- Very targeted questions to pull out specific details for an issue you're tracking down
- Broad questions to help you understand general aspects of your case materials
- Tactical questions to prepare for specific upcoming events, like depositions
- More "traditional" queries to identify sets of documents meeting specific criteria
Enabled search terms
The following terms are enabled by the Claude-Everlaw MCP connection:
- The Contents term
- The Binders term
- The Bates/Control # term
- Terms in the Metadata category of the Everlaw search page, such as email participants, custodians and, dates
Within Claude, you can write queries in natural language that reference these terms. Claude will translate them into Everlaw syntax and return the results.
Use your Project ID
As modeled in the example questions below, note the Project ID when you start a line of questioning. This tells Claude what project to search. To find your project ID:
-
In Everlaw, select the Help
button, then Contact Support.
This opens the Contact Support dialog. -
Your project number is the Project ID.
The project ID is also in the URL bar:
Subsequent questions within the same chat do not need to reference the project ID.
Understanding Claude vs. Everlaw AI
For users accustomed to using Everlaw search and Everlaw AI tools (such as Deep Dive) to search and synthesize documents, it's important to understand how the experience is different with Claude. There are two main differences we want to highlight:
- In Everlaw, the AI tools are intentionally confined to "the four corners of your document" and do not leverage outside information in responses. In contrast, Claude has access to knowledge outside your documents and brings additional context and information into its responses.
- Documents enrolled in Deep Dive, Everlaw's generative AI powered question and answer tool, have been ingested and indexed to be instantly searchable to generate responses to your natural language questions, even across millions of documents. Claude has not pre-ingested your Everlaw documents, and uses traditional keyword and metadata searches to understand your document contents.
Keep these differences in mind as you think through how you want to leverage this connection and plan your searches.
As with all AI-generated content, it's important to validate and verify the responses you receive within Claude.
Suggested questions and workflows
Below are some specific examples of the kinds of questions you can ask. They can help you think of questions that you might want to ask Claude about your Everlaw documents:
- Search and summarize a large document in your project: "In Project 2, retrieve and summarize the document with Bates Number #509574."
- Visualize and QA a small set of documents in your project: "In Project 2, visualize the "Top Docs" binder, grouped by year."
- QA and analyze your document set in your binders: "In Project 2, analyze the metadata in my binders and flag any data gaps."
- Tabulate results of documents meeting specific criteria: "In Project 2, find all the emails sent by Sanchez and containing the word 'meeting' or 'schedule'. Put them into a table that displays their date."
- Save time preparing for a deposition: "In Project 2, look in the Important Binder and filter it to documents authored by the deponent. Then build a document preparing me to conduct the deposition, including a line of questioning."
Tips for using Claude search
There is some nuance to how Claude builds its searches, especially if you are familiar with building them in Everlaw. Use these tips, and the examples, to make sure you get the exact results you're looking for:
-
Multi-word phrases: Claude does not always interpret quotation marks as a directive to search for an exact phrase. To make sure that multi-word phrases are searched exactly and not as words connected via an OR, we recommend that you either:
- Explicitly call it a phrase: Look for the exact phrase "stock market"
- Put the phrase in double quotes: ""stock market""
-
Proximity searches: Proximity searches typically do not work well without additional instruction. We recommend linking Everlaw's article about advanced content searching and requesting that Claude use the instructions there when doing a proximity search:
- Use the instructions for Proximity searches in this article (https://support.everlaw.com/hc/en-us/articles/210132673-Advanced-Content-Searches-Wildcard-Proximity-Fuzzy-Regular-Expression#ProximitySearches) to look for werewolf within 10 words of village
-
Time zone: Claude defaults to using the UTC time standard. To search for datetime metadata in a specific time zone, such as your project time zone, specify the time zone in your search:
- Search for all emails sent on May 11, 2024, using the Pacific Standard Time time zone.
-
Data size: When considering data sizes, Claude defaults to binary definitions of size (1KB = 1024B), but Everlaw uses decimal definitions (1KB = 1000B). When searching for data of specific sizes, tell Claude to use decimal sizing:
- Search for all documents greater than 1GB, in decimal size
-
Get links to document sets: If you want your search results to include a link to the results table of your search or binder, make sure you ask Claude to include it:
- For every answer that references an Everlaw search or binder, provide the Everlaw URL for it using the following template, using the Everlaw search ID (never the binder ID) to populate the numbers at the end of the URL: https://app.everlaw.com/<projectid>/search.do#id=<searchID>
Claude Project instructions example
More advanced Claude users can use Projects to provide Clade with specific instructions on how to search Everlaw.
Here is example text a Claude user could paste into Claude Project instructions to optimize Claude’s Everlaw searches [replace all caps in brackets with appropriate values]:
- Use Everlaw project [PROJECT ID] for all questions.
- Assume the [YOUR EVERLAW PROJECT TIMEZONE] timezone for all date/time searches.
- Assume decimal definitions for document sizes (e.g. KB, MB, GB), not binary.
- For contents searches, assume all text within a set of quotation marks is a phrase, not a set of individual terms.
-
To run a contents proximity search, assume the syntax in the following examples:
- -for single word proximity where "werewolf" is within 10 words of "village": "Werewolf village"~10
- -for phrase proximity where "flight from" is within 20 words of "werewolf village": "("flight from")("Werewolf village")"~20
- -when order matters, where "werewolf" is within 10 words before "village": "Werewolf village"~~10
- For phrase proximity searches, do not add escape characters when you send the request to Everlaw with JSON string encoding.
- For every answer that references an Everlaw search or binder, provide the Everlaw URL for it using the following template, using the Everlaw search ID (never the binder ID) to populate the numbers at the end of the URL: https://app.everlaw.com/[PROJECTID]/search.do#id=<SEARCHID>
Capabilities and limitations
The below sections are non-exhaustive lists of the capabilities and limitations of searching Everlaw using Claude. We're putting them here to give you a sense of the kinds of analysis you can and can't do, which might help you define the types of searching and analysis you're interested in using Claude for.
Capabilities
Claude can:
-
Search your document text, document metadata, Bates/Control numbers, and the binders you have access to
Note: Claude's access to full document text, as opposed to keyword search results returned by Everlaw, is based on your team's settings. See the All network egress section in our setup article for more details. - Identify the number of coded, highlighted, and redacted documents
- Create and display a markdown table of results (analogous to an Everlaw results table CSV export), with links to open each document in the Everlaw review window
- Create visualizations, such as histograms, or dynamic interactive dashboards, to show the distribution of results over specified criteria/metrics
- Have a "conversation" that maintains a chat history, and can reference previous questions/responses to build on your analysis
- Compare two projects within a conversation, such as asking "Which project has more documents that mention 'energy?' Project 2 or project 38?"
Limitations
At this time, Claude cannot:
- Search details about review work such as the specific codes or ratings applied to documents, or who applied redactions or highlights
- Search the contents of stored translations
- Consistently interpret or correct typos in your contents queries. If you ask it to search for documents with the word "enegy" instead of "energy", it will likely look for "enegy."
- View or analyze images/non-text content
- Take actions on a document that are reflected within Everlaw, such as applying redactions or codes
- Access any information about project users or analytics
- Connect to Storybuilder
- Use Deep Dive to answer a question
- Claude might struggle to answer questions that involve many documents, or specific kinds of metadata
- Claude might struggle to build visualizations of more than 4000 documents