Using the chat

Guide · Users · Applies to 5.0.196

The web chat at https://<host>/chat is where most people use Eldric. This page explains what happens when you send a message, how to run a tool yourself, and what the cards and notes under a reply mean.

Tools Eldric chooses

From 5.0.187 Eldric chooses its tools by itself: it searches your knowledge bases, looks something up in memory, searches the web or reads a device when the question needs it. It only uses tools your role allows. The tool calls of a reply stay with the message, also after a reload and in an export.

Under each reply a line says which model and service answered. Administrators also see which node it was.

From 5.0.196 the model picker offers only models that can chat; models for search and ranking are left out. If the cluster has no chat model at all, it says: "The models on this cluster are for search and ranking, not for chat. Ask an administrator to add a chat model."

Running a tool yourself: "/"

Type / in the message box to see the commands. From 5.0.192 the list includes every tool your role may use. A tool you pick runs straight away, without the model:

• A tool that takes one text takes the text after its name: /websearch eldric release notes.
• Other tools take name=value pairs (quote values with spaces) or a JSON object. If something is missing, the chat shows the tool's usage.
• The hint under the message box lists the tools you can use. It also names tools you were granted that are not available to your role on this server.

Some useful commands: /new starts a conversation, /export saves it, /memory opens your memory, /kb your knowledge bases, /remember and /forget change memory (see Memory), /help lists everything.

Cards under a reply

• Memory cards show what the server did when Eldric stored, changed, recalled or removed something: Memory updated with the old and new value, Stored in memory, Removed from memory, or why nothing happened. A recall lists each fact with its condition, whom it is about and when it was learned. See Memory.
• Methods: when Eldric offers a method that worked before, the card lets you choose it and asks afterwards whether it worked.
• Error cards explain why a reply did not come, for example that the conversation is too long for the model (see below) or that the model is not available to you.

Code and files in replies

• Every code block in a reply is also saved as a file you can find again under /artifacts. From 5.0.191 the file is named after what it holds, for example project-summary.json, instead of code-1.json.
• Code blocks show the characters themselves, and an SVG block shows the picture. A wide preview scrolls inside its card.

Web search

Web search is built in; it needs no plugin. If your administrator has stored a Brave Search key, Eldric asks Brave first and uses other search engines as a fallback. Each answer says which engine answered. /websearch <query> searches without the model.

Long conversations

From 5.0.190 the chat knows how much text each model can take. When a conversation grows too long, the chat leaves out the oldest messages and says so under the reply. If even that is not enough, a card shows the size of the conversation and the model's limit. Start a new conversation or choose a model with a longer context.

Deleting and exporting

• Deleting a conversation removes it from your device and from the cluster, and it will not be read into Eldric's memory afterwards. Whether what Eldric already learned from it is forgotten depends on your administrator's setting; the confirmation tells you. See Your data.
• /clear empties the current conversation, also on the server.
• /export saves the conversation as Markdown, HTML, JSON or text.

Working with your documents is described in Your documents.