Chat

Chat is a private conversation with a model running on Ferrith's infrastructure, with your documents alongside it. Conversations are private and cannot be read by admins. The only time they become visible is within the data package provided when the owner performs a full export of data from Ferrith.

Start a conversation

1. Click + New in the sidebar.
2. Pick a model from the dropdown if more than one is offered.
3. Type and send. Replies stream in as they're written; each reply shows its token counts underneath, and a copy button for taking the text elsewhere.

The sidebar lists your conversations, most recent first, with filter chips (All / Chat / Agent / Workflow) once you're using the other modes below.

Attach a document

Attach a document when you want the conversation grounded in it.

1. Use the paperclip in the composer and choose a file, PDF, Word (DOCX), Markdown, or plain text that has already been uploaded to the library.
2. New documents can be atached to Workflow chats. Wait for the Processing chip to clear as a newly uploaded document needs to be read and indexed, which takes seconds to a couple of minutes. You can't send against it until it's ready.
3. Ask your questions. Small documents are read in full; large ones are searched, and the answer draws on the most relevant passages.

Uploads are accepted up to your workspace's size and page limits. The uploader shows the current figures (typically up to 50 MB, and up to 300 pages for a PDF; split anything longer). A conversation can carry several attachments.

Attach a folder

Attach a whole folder instead of a single file when the answer could be anywhere in a set, a matter folder, a policy library.

• The folder's current contents are used, so a document added to the folder later is included in later questions.
• Answers tell you honestly what they saw: a small document is marked as read in full, a large one as excerpts only — and the assistant is instructed never to claim something is absent from a document it only sampled.
• Naming a document in your question ("in the Master Services Agreement, …") focuses the search on that document.
• Multi-doucment or folder based chat is better suited to analysis grids

Citations and the document viewer

When a conversation has documents attached, replies cite their sources as numbered references. Click one and the document opens beside the chat at the cited page, with the quoted text highlighted.

If a highlight can't be located (scanned pages, heavy formatting), the viewer still opens at the cited page.

Download an answer

Under any assistant reply, use the download control to save it as a PDF or Word document, useful when the reply is the deliverable. You can name the file before saving.

Your daily allowance

Model usage is metered against your workspace's daily allowance. As you approach your share of it, a small chip appears near the composer; if a limit is reached, sending pauses and the message tells you exactly when it resumes. Admins can see the workspace-wide picture on the workspace details page.

Long conversations

Models can only read so much at once. In a very long conversation, the oldest messages stop being sent with each new question, and a "earlier messages aren't being sent from here" divider appears at the point where that began. Nothing is deleted, the full conversation stays on screen, but the model no longer sees what's above the divider. For a new topic, a new conversation always gives the best answers.

Chat with an agent

The composer's Agent tab lists your workspace's agents, assistants with their own instructions and knowledge base. Pick one and the conversation runs as that agent: its greeting appears first, and its knowledge base informs every answer. The model choice is the agent's, shown as a read-only chip.

Start a workflow from chat

The Workflow tab lists the workflows you can run. Pick one and the workflow's intake assistant opens the conversation and it asks for whatever the workflow needs. When it has everything, a start control appears on its reply; starting hands the conversation's result to the workflow and takes you to the live run. The conversation locks at that point (it's now part of the run's record), and the run picks up from what was gathered, you can follow every step as it executes.

A note on AI-generated content

Answers are generated, and generated content can be wrong. Citations exist so you can check claims against sources in one click; for anything you'll rely on you should validate it first. Documents produced by chat or workflows are drafts for review, not finished advice.