AI chat¶

This page is for attorneys and paralegals who want to ask an AI model about a matter. It covers starting a conversation, what the AI is and is not given, what it can add to the matter for you, and how to check what it tells you. It also covers the smaller chat on an intake.

Where AI chat appears¶

Kosmos has no AI of its own, and AI is optional. An administrator connects it to Anthropic (the Claude models), to Google (the Gemini models) or to both, under Settings → Integrations (see AI). Until one is connected, none of the places above appear: there is no AI tab on matters and no Chat on intakes.

Start a conversation on a matter¶

1. Open the matter, click Case, then AI. The tab lists the matter's conversations.
2. Click the plus button at the left of the toolbar. New Conversation opens.
3. Choose a Mode and a Model, type a Name and click Submit.

The conversation opens in a new browser tab, with the matter, the name and the model across the top. It joins the list on the AI tab when you send your first question.

Model lists the Claude and Gemini models of the providers your server is connected to. The mode and the model are fixed once the conversation exists. An Agentic badge marks those conversations in the list and in the chat window.

Good to know:

• A conversation belongs to the matter, not to you. Everyone who can open the matter can read it and add to it, and each question shows who asked it.

Ask a question¶

1. Type in Ask about this matter... and press Enter. Shift+Enter starts a new line. For a long question, click the expand button beside the box: Compose Prompt opens with formatting buttons, and Send to AI sends it.
2. Watch the status under your question while the AI works.
3. Read the reply. Ask a follow-up in the same box: the AI is given the conversation so far.

In a Classic conversation the status is one line that changes as the work moves on ("Building context...", "Connecting to AI...", "Thinking..." or "Generating response...", "Verifying citations..."), with the seconds elapsed and a list of what was gathered and selected. In an Agentic conversation each step is a row ("Read", "Searched"), and the bar above the message box shows the time, the size of the request and how many lookups are used, for example "7/40 tools".

To stop, click the x button beside the status or press Esc. Stopping also removes the question you asked.

Above the reply, Activity (Classic) or Agent run (Agentic) lists the steps behind it. The bar above the message box shows the size of the conversation and an estimated cost for the last reply and for the whole conversation. History, at the left, lists the questions asked: click one to jump to it. The copy button on a reply copies it with its formatting. On a question, the delete button removes that question and its reply, and the scissors button moves that question and everything after it into a new conversation.

Good to know:

• Keep the chat tab open until the reply appears. The work carries on if you close the tab, and Resume shows it again, but a reply that has been finished for more than ten minutes without being shown is lost: your question stays, with no answer and no message. Anything the AI added to the matter in that reply is still there.
• "This request was interrupted before it finished (the server restarted mid-run). Please re-send your message." means the server was restarted while the AI was working. Send the question again.
• A reply that starts "Error: Unable to get response." gives the reason after it. Send the question again, or start a conversation with another model.

What the AI is given¶

A chat on a matter works from that matter's record. Nothing from another matter is given to it. Two things come from outside the matter: the firm's library notes (see Notes), and the names and titles of the firm's users, with your own name and email address.

What it is given also depends on who asks. Time entries, with their rates and fees, are given for every user who can open the matter, as the matter's Activity tab shows them. Invoices and the matter's rates are given only if you have the Financial permission, and published case law is searched only if you have the Research permission (see Permissions). Ask your administrator.

In a Classic conversation¶

Given with every question:

• the matter's name, status, work status, practice area, description, open date and client;
• its contacts (name, role, company and email), its witnesses and its proceedings;
• every highlight and every timeline fact (see Timeline, witnesses and highlights);
• up to 20 tasks (open ones first), the next 15 events and the last 5 past events;
• every time entry, with its hours, rate, amount and invoice status, and the settlement log;
• in full, each document, saved case and conversation whose AI setting is Always.

Chosen for each question:

• documents set to Auto whose text has been extracted (see Documents);
• the matter's notes and its synced email threads, with the text of their attachments (see Email);
• saved cases (the Case Law view of the AI tab, see Case law) and other conversations set to Auto;
• invoices, only when the question is about billing and you have the Financial permission;
• library notes, only when one bears on the question.

If these items come to less than about 30,000 words, and there are no invoices or library notes to consider, all of them are loaded. Otherwise a first AI step is shown your question and a list of the items (name, category, date, a short summary, size and importance) and picks the ones to load in full. The rest are given by name only, so the AI can tell you that a document exists which it has not read.

In an Agentic conversation¶

The AI starts with the matter's details, contacts, witnesses and proceedings, the highlights and timeline when they are short, and an index with one line for each document, email thread, note, saved case, earlier conversation and library note. From there it opens what it needs, up to 40 lookups for each question. It can also read the matter's tasks, events, settlement log, and time, expense and flat-fee entries. With the Financial permission, the index lists the matter's invoices as well, and the AI can read them and the matter's rates. With the Research permission, it can search published case law on CourtListener. Without it, the AI works from the cases saved on the matter. What it has read stays with the conversation for your later questions.

The AI setting¶

Documents, saved cases and conversations each have an AI column.

What it is never given¶

• Records of any other matter.
• The matter's ledger and the client's trust records.
• Invoices and the matter's rates, unless you have the Financial permission.
• Anything set to Never, and the contents of a document whose text has not been extracted.

Good to know:

• Permissions are checked when a question is asked, not when a conversation is read. A reply to a colleague who has the Financial permission can quote invoices and rates, and everyone who can open the matter can read that conversation.
• In a Classic conversation, a follow-up sent within ten minutes works from the items chosen for the earlier question, unless the case file changed in between. If the AI says it has not seen a document, ask again later or set the document's AI to Always.

Have the AI add to the matter¶

The AI answers in prose unless you tell it to record something. Use the word "timeline", "witness" or "note" in the request: one that leaves the word out may be answered in prose only.

A matter chat does not create tasks. Add them yourself (see Tasks). To have the AI edit a document you are drafting, link the draft to the conversation with the pen button beside the message box (see Drafts).

Good to know:

• Nothing is confirmed first. The items exist as soon as the reply appears, so read each line and correct mistakes on the tab where the item landed.
• The AI adds to the end of a note unless you tell it to rewrite or replace the note. A rewrite replaces the whole note, and a firm library note can be rewritten this way from any matter.
• A line saying the assistant described a note write without performing one means no note was changed. Ask again.

Check sources and citations¶

A reply links the documents and highlights it relies on. The link text is the document's name or the highlight's citation. Click it to open the document, at the highlighted passage when the link is to a highlight. The link opens in the chat's own tab: use your browser's command to open it in a new tab if you want to keep the chat on screen.

Each case citation in a reply has a small badge after it. It spins while Kosmos looks the case up on CourtListener and reads the opinion against what the AI said about it, then shows a check mark (the case supports the point), a minus (it supports part of it), a warning triangle (it contradicts it) or a question mark (unclear). Click the badge to open Citation Check, with what the AI used the case for, the rationale, a quote from the opinion and a link to the opinion. The bar above the message box shows "checking cites", then "cites ok" or the number that could not be found.

Manage conversations¶

The AI tab lists each conversation with its Date, name, Model, number of Messages, Participants, Last Activity and AI setting. Type in Filter conversations . . . to match names and the text of messages, or choose a person under All participants.

Click a conversation's name for its menu:

To act on several at once, tick their boxes: the toolbar then offers AI (to set them to Auto, Always or Never) and a trash button. A conversation's AI setting decides whether later chats on the matter are given it, as described above.

Summaries¶

After each reply, Kosmos writes a summary of about 100 words of the conversation so far. It appears as Conversation Summary at the top of the chat window the next time you open it, and it is what describes the conversation when a later chat chooses what to load.

The intake chat¶

Chat on an intake opens Intake Chat. It uses Gemini when it is connected and Claude otherwise, and keeps the same one for the life of the chat. It is given that intake's details, notes and assessment and can change the intake's fields when you tell it to. An intake has one chat, shared by everyone with the Intakes permission. The save button ends it and posts a summary of its conclusions to the intake's notes. See Assessment.

How long chats are kept¶

Conversations on a matter stay until someone deletes them, with one exception: once a matter has been Closed for a retention period (180 days unless your administrator has changed it), its AI conversations are deleted. See Close a matter. Timeline facts, witnesses, notes and saved cases that the AI added are part of the matter and are not deleted with the chats.

What leaves the firm¶

Each question is sent, with the parts of the matter described above, to the provider of the model you chose: Anthropic for a Claude model, Google for a Gemini model. Whichever model you choose, the supporting work (choosing what to load, writing summaries and checking citations) goes to Google's Gemini when it is connected, and to Anthropic when only Anthropic is. The text of each reply is sent to CourtListener to look up the cases it cites, and so are the case law searches of an Agentic conversation. Whether that is acceptable for a given client is a decision for the firm: AI providers and research lists exactly what is sent.