Chat

Use Chat to talk to a model, compare answers, work with files, speak instead of type, or ask an agent to use tools. SoAI saves each transcript with the models, mode, prompts, generation settings, workspace folder, tools, skills, and Knowledge used in that conversation. Knowledge is an indexed document collection that the assistant can search. Open the conversation later and those settings return with it. See Conversation context for what SoAI stores, sends to the model, and compacts as the transcript grows.

Before Chat appears

SoAI setup in the first-run wizard can prepare a model and open Chat with it. On a fresh installation with no selectable conversation model, SoAI leaves Chat and Automation out of the main sidebar. To add a model yourself:

  1. In the SoAI WebUI, open Models and select Add model.
  2. Use Download model for a local chat model. If it needs a local backend first, follow Install and manage plugins. For a remote model, select Add provider and use the bundled External plugin.

Chat appears after the catalog contains a selectable conversation model, meaning the model is loaded or available and is not embedding-only. A provider configuration alone does not count; discovery must add one of its models to the catalog. See Models catalog for model discovery, downloads, and availability.

When Chat offers Finish SoAI setup, select it to continue the remaining setup steps.

An embedding-only model does not unlock Chat because it cannot answer conversation prompts. It can still supply the vectors used by Knowledge. After this first-run gate opens, that browser remembers it; Chat can remain in the sidebar if the model later becomes unavailable, but you must select or restore a selectable conversation model before you can send a message.

Conditional controls

Chat shows actions when they apply. Visibility and availability depend on the model, conversation state, screen width, browser support, account access, and Appearance settings.

Control When it appears or becomes available
Approval, question, and secret panels These appear only when an agent has paused for that specific response.
Compact and Export conversation These actions require at least one message. They are unavailable while a response or agent turn is running.
Controls in automation or messaging conversations The Automation or Messaging configuration that created the conversation can hide or lock its execution settings. Change them at the source rather than in Chat.
Header and composer actions On a narrow screen, actions such as Favorite, Tools, the plan toggle, and some composer controls move under More actions or into the compact header.
Message actions Hover over a message or focus it with the keyboard. The message role and current state decide whether actions such as Edit, Resend message, Regenerate, Speak, or Message info apply. History-changing actions stay unavailable during a turn.
Model selector in the header The sidebar model selector is the main control while the sidebar is open. The header selector appears when the sidebar is closed or Chat is in a detached window.
Previous variant and Next variant These appear only on a comparison turn that has more than one model reply.
Saved prompts, character map, sketch, token counter, voice, attachments, and camera Turn each action on under Chat configuration > Appearance > Input actions. The Upload and Scan panes follow the attachment and camera choices. Dictation can remain disabled when the page or browser cannot record audio or permission is missing. Voice Call also requires a selected conversation model. Scan needs browser camera support and permission.
Toggle plan bar and Toggle todo bar These appear after the current conversation has a plan or todo list.

Page layout

Area What you can do
Conversation sidebar Select a model, start a conversation, search conversations, move favorites to the top, open a conversation, and use selection mode for batch actions.
Conversation header Rename the conversation, change mode, compact context, export, open Chat configuration, mark the conversation as a favorite, manage tools, show a plan or todo list, and start another conversation.
Timeline Read messages, reasoning and activity states, tool calls, plans, todos, attachments, approval requests, and errors. Older or newer windows of a long transcript load as you move through it.
Interaction previews Review queued prompts, approve tools, answer agent questions, provide requested secrets, monitor voice recording, and inspect attachments before sending.
Composer Write a prompt, attach context, use saved prompts, dictate, start a voice call, inspect token use, and send, stop, steer, or queue input.

Chat can hide its conversation header and empty composer while you scroll up through earlier messages. Scroll down to reveal them, or move the pointer to the top edge for the header and the bottom edge for the composer. Typing brings back the composer, and a running response keeps its Stop control available. Turn off Auto hide chat header or Auto hide chat composer under Chat configuration > Appearance to keep them visible. With the conversation sidebar collapsed, rest the pointer on its toggle button to open it over the page.

Start a conversation

  1. Select New chat in the conversation sidebar, or New conversation in the header.
  2. Select the current model name at the top of the sidebar, then choose an available model. You can also use the model selector in the header or empty conversation. Search narrows the list, and availability labels show which models are ready to use.
  3. Select the mode label in the header and choose Chat, Plan, or Execute.
  4. Write a message. Add files or configure tools and Knowledge if you need them, then send it.

Defaults and titles

A new conversation begins with an empty transcript. Use last settings for new chats is enabled by default and reuses your last model, tools, and Knowledge settings. Turn it off to use the configured defaults. Change it under Chat configuration > General, then select Save. SoAI saves a new conversation when you first send or use a feature that needs a saved chat.

Turn on Auto-rename conversations under Chat configuration > Appearance to generate a title after the first primary-model reply. Select the header title to rename a conversation yourself; a saved name takes precedence over automatic naming.

Model requirements

Select an available model before sending. If a saved model disappears from the catalog, SoAI keeps its name visible and marks it (missing); choose another model or fix that backend before you continue. Tools need a model with tool-calling support. Image input also needs a model and backend that accept images. See Models catalog for model setup and capability details.

Choose the conversation mode

Select the current Chat, Plan, or Execute label in the header to choose how the assistant should work. The same control appears in an empty conversation. You can also press Shift+Tab to cycle through the modes. Each mode has its own tool selection.

Mode Use it for Tool behavior
Chat Questions, writing, analysis, and ordinary conversation. Tools are optional. When you turn them on, SoAI uses the conversation's Chat tool selection.
Plan Explore a task, clarify requirements, and write a plan without making changes. Plan keeps tools enabled but blocks destructive tools and tools without enough safety information.
Execute Carrying out a multi-step task with progress tracking and tool use. Execute keeps tools enabled and uses the Execute tool selection. Those tools can change files, run commands, control a browser, or affect external services.

Plan and Execute need a model with tool-calling support and at least one tool selected for that mode. Changing the mode cannot grant access to disabled tools or tools outside your account permissions. Finish or stop the current turn before changing its mode or tools.

Use one model or compare several

Open the model selector and select the primary model. To compare answers, select Compare models, then choose models in the slots labeled Model 2 through Model 5. You can use one primary model and up to four comparison models in the same turn. Change or remove a slot while the conversation is idle.

SoAI sends the prompt to every selected model and groups the replies into one comparison turn. Use Previous variant and Next variant to read each reply. Every variant shares the earlier transcript and stays in the same conversation. The carousel changes the visible reply. Later prompts use the primary model's reply as conversation history, even when another variant remains visible.

You cannot steer a comparison because several models are answering. If SoAI can accept another prompt, the composer shows Queue; otherwise wait or select Stop.

Write, send, steer, and queue prompts

The composer accepts multiline text and direct attachments. On an empty conversation, suggestion cards can insert a starting prompt. Select Saved prompts beside the composer to insert an entry from your prompt library. Select Manage prompts in that picker to create or edit entries on the Prompts page. Select Open character map, choose one or more characters, then select Insert when you need characters that are inconvenient to type. Select Open sketch to draw with a mouse, pen, or touch, then select Attach to add the drawing to the composer as an image. If an action is missing, turn it on under Chat configuration > Appearance > Input actions, then select Save.

The character map loads more results as you scroll. Within its grid, use the arrow keys to move between characters, Home or End to reach the first or last result, and Page Up or Page Down to move through larger groups. Search and the selected Unicode range apply to the full catalog.

Auto hide chat composer is enabled by default under Chat configuration > Appearance. When the composer is empty and unfocused, it folds away as you scroll up through earlier messages. Scroll down, type, or move the pointer to the bottom of Chat to reveal it. The Stop button remains visible while a response runs, and agent questions or approval requests bring the composer back when your attention is needed. Turn the setting off if you prefer a fixed composer.

SoAI saves the composer text and ready direct attachments as a draft for that conversation. Switch conversations or reload the page and the draft returns. Drafts also sync between signed-in browsers. If another browser changes the same draft while you are typing, SoAI keeps your local text and shows a conflict instead of replacing it.

The primary composer action changes with the state of the current turn:

  • Send starts a new turn when the conversation is idle.
  • Stop ends the active response or agent turn when the composer is empty.
  • Steer gives new text and ready attachments to the running turn. An agent turn takes the steer at its next step, such as after a tool call. While the model is writing its reply, the steer waits until the reply finishes. Select Stop or press Esc to end the reply and send the steer immediately; the stopped reply then notes that it continued below with your new message.
  • Queue stores the text and ready attachments for the next available turn. Pending queue items appear above the composer and can be cancelled before they start.
  • Resume queued prompts starts paused queue items when the conversation is idle and the composer is empty.

The button label tells you what will happen. An approval, agent question, secret request, comparison, or existing queue item can limit the next action. New text will steer only when the button says Steer.

Stop pauses the prompts waiting in the queue, so they do not start after the stopped response. Each paused item shows Paused above the composer and can still be cancelled. A message you send afterward runs ahead of the paused items, and they continue once a response completes without being stopped. To continue them without sending another message, leave the composer empty and select Resume queued prompts.

Keyboard controls

Shortcut Action
Ctrl+Enter or Cmd+Enter Send when idle, stop when running, or save an active message edit.
Shift+Enter Insert a new line.
Tab Send when idle. During an active turn that accepts more input, queue the composer text and attachments.
Shift+Tab Cycle Chat, Plan, and Execute mode.
Esc Stop an active turn, cancel an agent question, or exit conversation selection mode.
ArrowUp at the start of the composer Move backward through your saved prompt history.
ArrowDown at the end of the composer Move forward through your saved prompt history.

Hold shortcuts

Press and hold these controls to open another action.

Control Hold action
Start voice recording Record while held, then release to transcribe and send. See Using speech in Chat for cancellation and draft review.
Token counter Open Chat configuration on Parameters with the context length in view.
Chat configuration Open Presets.
Toggle tools Open Tools settings for this conversation.

Attach files, folders, and media

Add attachments opens Upload, Scan, Workspace, SoAI link, and Knowledge. One message can mix files from several of these sources. Direct files must finish processing before you send; linked workspace items remain verified references, while Knowledge builds a searchable conversation collection.

See Attachments and file context for every WebUI attachment path, including folder uploads, Global Search and Files links, SoAI links pasted into the composer, images, audio, video, document parsing, limits, and conditional controls.

Build indexed Knowledge for a conversation

Use Knowledge for documents that the assistant should search across several prompts. Knowledge needs an embedding-capable model to index and search the collection, plus a conversation model with tool-calling support to use the results while answering. One model can fill both roles when it supports both capabilities.

Your account needs RAG access (RAG_USE) to use Knowledge. Importing from Workspace also needs Files reading (FILE_EXPLORER_READ). The default policy grants both actions to Administrator and Standard user accounts. An Administrator with ACL administration (ACL_ADMIN) selects the Settings button at the right of the header, chooses Settings, then opens ACL to change them.

Prepare an embedding model

Auto (embedding) chooses from embedding models already available to SoAI. It does not download a model or add embedding support to a chat model. Prepare a local or remote embedding model before you add documents.

Source WebUI setup
Local Llama.cpp Embeddings plugin Open Plugins, find Llama.cpp Embeddings, and select Install backend if its backend is not installed. Choose a backend variant if SoAI asks, then select Start installation. Next, open Models, select Add model > Download model, choose Llama.cpp Embeddings as the plugin, and use Model search to find a GGUF embedding model. Choose a model and quantization, then select Download model. The Bundled plugins reference describes the dedicated local Embedding runtime.
Remote provider through External Open Models, select Add model > Add provider, choose External, and enter the provider name, OpenAI-compatible API URL, API key, and optional Models filter. Leave the filter blank or include the embedding model's ID or a matching pattern; an unmatched filter keeps that model out of SoAI's catalog. The API URL is the provider's OpenAI-compatible base URL, such as https://api.example.com/v1, rather than a direct /models or /embeddings URL. Select Add provider. Choose a model that the provider documents as embedding-capable; the provider must support the OpenAI-compatible embeddings endpoint. No local backend installation is needed for External.

After the download or provider discovery finishes and SoAI refreshes the model catalog, return to Chat configuration > Knowledge. Check that the model appears under Embedding model. If the list says No embedding models available, the current catalog has no available model marked with embedding support. Select the model directly, or use Auto (embedding) once at least one suitable model is listed.

If Auto cannot find a suitable embedding model, indexing stops with guidance to select or prepare one. It does not substitute an ordinary chat model. Select a model explicitly when it supports embeddings but is not identified as an embedding model in the catalog.

Add and index documents

Add and manage the collection from Add attachments > Knowledge:

  1. Select Upload file, Upload folder, or Load from Workspace.
  2. Follow the indexing status. A document can move through queued, parsing, chunking, and embedding before it is ready. SoAI marks unreadable files and errors in the same list.
  3. After a document finishes processing, select Delete beside it to remove it from this conversation's collection.
  4. Select Reindex after changing the embedding model or chunking settings.

Reindex requires at least one successfully indexed document. If the collection has none, upload and index a document first, then use Reindex when you need to rebuild that existing collection.

Change retrieval settings under Chat configuration > Knowledge. Turn on Enable RAG, select an embedding model or Auto (embedding), choose the retrieval and chunking values, then select Save. These settings belong to the open conversation.

When RAG is enabled, the collection has a document, the Knowledge tools are available, and the conversation model supports tool calling, SoAI keeps those tools enabled for the conversation. The ordinary tool toggle cannot remove them until you disable indexed Knowledge. Without tool-calling support, the conversation model cannot search or list the collection even when its documents are ready.

Knowledge control Available values Effect
Chunk overlap 0 to 500 Repeats content between adjacent chunks, in the unit the chosen strategy counts. Paragraph ignores it. For token based and fixed size, overlap must be smaller than chunk size.
Chunk size 100 to 4,000 Sets the target chunk size. Token based counts tokens. Fixed size, paragraph, and semantic count characters.
Chunking strategy Token based, fixed size, paragraph, semantic Splits a document before embedding, which decides what a search can return.
Retrieval strategy Similarity, MMR, Hybrid Ranks the passages a search returns.
Similarity threshold 0 to 1 Filters weak matches.
Top K 1 to 50 Limits how many chunks a search can return.

Direct Upload attachments belong to one message, while Knowledge documents belong to the conversation's indexed collection. Use Upload when a file belongs with the current prompt, and use Knowledge when you want the assistant to search a collection across later prompts. The values above are what this form accepts; Knowledge and document collections (RAG) covers which chunking and retrieval strategy suits a source, what the indexing stages mean, and when a collection has to be reindexed.

Use reusable skills

With MCP usage (MCP_USE), open Chat configuration > Skills, select Add skill, and import a SKILL.md file or one-skill ZIP. Read it with View, enable it, then select Save. Importing a skill does not enable it for the conversation.

With tools available, the assistant can load an enabled skill when its description matches the task. To request it directly, write Use $review-notes to review this draft. or start your message with $review-notes. Direct requests can supply instructions with tools disabled, but actions still need their normal tools and permissions. New chats start with no skills selected. See Skills for loading options, presets, supporting files, and limits, or Creating a skill to write your own.

Configure and approve tools

Use the Tools control in the header to turn tools on or off for a Chat conversation. Plan and Execute keep tools on. To choose specific tools, open Chat configuration > Tools, search the list, expand a server, and select tools for Chat mode, Plan mode, or Execute mode. Select Save when you finish.

Your account needs MCP usage (MCP_USE) to open the tool picker. The default policy grants this action to Administrator and Standard user accounts. An Administrator with ACL administration (ACL_ADMIN) selects the Settings button at the right of the header, chooses Settings, then opens ACL to change it.

SoAI locks the Tools control during an active turn. Indexed Knowledge also keeps its required tools on; turn off Enable RAG under Chat configuration > Knowledge before you try to disable them.

To edit the starting lists for later conversations, select Select under Default tools (new chats). This dialog is another view of the staged tool lists for the open conversation, so its changes also appear in the main tool picker. Close the dialog, review the selections, and select Save to update both the open conversation and your defaults for later chats.

The list contains the built-in tools and any tools from remote MCP servers available to your account. Chat can use web search, browser control, file editing, shell commands, memory, or image generation only when the matching tool appears in the list and you select it for the current mode. See Agent tools overview and MCP overview for the catalogs and server setup. When a saved tool or MCP server is no longer available, the tool settings list it as unavailable and keep it selected for when it returns.

Approval prompts

Turn on Require tool approval under Chat configuration > Tools to pause before a destructive tool, a tool flagged as requiring approval, or a tool with missing or invalid destructive metadata. The approval panel names the tool and shows its proposed input. Select Approve or Deny.

Select Remember for this tool when you trust later calls covered by the same saved permission. For most tools, that permission applies by tool name across your account's conversations, regardless of the arguments in a later call. Browser credential autofill is the exception: SoAI limits permission to the destination site. If it cannot determine the site, approve without Remember to allow the current call once.

Plan mode still blocks destructive tools and tools with missing or invalid destructive metadata. An Administrator selects the Settings button at the right of the header, chooses Settings, then opens Reset > Tool approvals to clear saved approvals.

Interactive agent checkpoints

Agent turns can pause for three user-facing interactions:

  • Tool approval asks whether a proposed call may run.
  • Question asks for one or more answers needed to continue the task.
  • Secret request asks for a credential and can offer to save it in the Secrets Vault with a label.

Answer the panel above the composer to continue. If you cancel it or let it expire, SoAI marks that interaction as cancelled, timed out, or failed. The agent does not receive a successful tool result.

Follow plans, todos, tools, and subagents

Plan and Execute turns can expose a plan or todo panel above the timeline. After the turn creates one, the matching Toggle plan bar or Toggle todo bar control appears in the header. Use it to show or hide the panel. The same progress updates remain in the transcript.

In Plan mode, the agent can place its plan in the Plan widget and leave a short summary in chat. Select View plan to review it, then Execute plan when you want the agent to carry it out.

Tool activity appears inside the assistant turn. A row can show the tool name, state, input, result, duration, URL, code diff, media preview, or error. Expand a row for details. Select View full output when a long result offers it. A running shell command has its own Stop command action.

When an agent delegates work, subagent activity appears in the same timeline with its status, timing, and nested activity. Tool results stay with the assistant turn that produced them.

Read and act on messages

Turn Render rich text (Markdown) on or off under Chat configuration > Appearance. Rich text gives code blocks a copy action and makes Markdown table headers sortable. Plain-text mode shows the original message text. If Markdown rendering fails, SoAI keeps the message and shows safe plain text.

Under Chat configuration > Appearance, turn on Inline multimedia previews to show supported media links and file references inside the timeline. Show activities chooses whether activity details begin expanded or collapsed. While a response is running, select its loading activity to show or hide the individual activities and their details. The collapsed loading activity continues to show what the model is doing in real time. Turn off Show activities elapsed time to stop updating the timer while an activity runs; the final duration still appears after the activity ends. Select Save after changing these settings. Hover over a message, or focus it with the keyboard, to show the actions that apply to it.

Message action What it does
Copy Copies the message content. Code blocks also have a dedicated copy control.
Delete Replaces the message with a temporary deleted placeholder. Choose Undo during the short undo window, or select Delete again to commit before the timer expires.
Edit Replaces a user message, removes every later message, and requests a fresh reply from that point. Direct attachments on the edited message can also be removed before saving.
Message info Shows available model, timestamp, prompt and completion tokens, total tokens, generation speed, latency, finish reason, and compaction statistics.
Regenerate Replaces an assistant reply and removes messages that followed that reply.
Resend message Resubmits a selected user message and removes every later message before generating the new continuation.
Speak Reads an assistant message using the selected or automatic text-to-speech model.

Edit, resend, and regenerate rewrite the continuation of the transcript. Export or clone the conversation first if you need to preserve the later turns. SoAI disables history-changing message actions while the conversation runs.

Use the token counter and compact context

Open Chat configuration > Appearance > Input actions, turn on Token counter, and select Save. The counter beside the composer shows estimated or measured prompt use and the model's context window when SoAI knows it. Select the counter to switch between token use and generation speed, or hold it to open Parameters with the context length in view. Its color changes as the prompt approaches the limit. Prompt use includes the enabled SoAI system prompt and its runtime context.

Context compaction

Before an agent request, SoAI may compact older history when the full prompt no longer fits the available prompt budget. To start compaction yourself, wait until the conversation has at least one message and is idle, then select Compact in the header. SoAI uses the selected conversation model, summarizes older context, and adds a visible boundary to the transcript. Later turns use that summary in place of the earlier detail. If the summary would not be smaller than the context it replaces, SoAI keeps the existing history and reports that nothing was compacted.

Automatic compaction also applies to messaging conversations. It lets a conversation continue without a fixed message-count limit. The server may replace older details with a summary instead of sending them to the model word for word. Each messaging account can use a different context length and therefore have its own effective compaction limit. Select the Settings button at the right of the header, choose Settings, select Messaging, then use Edit for the account. Under Chat execution, select Edit parameters and set Context length (tokens) under Model settings. Clear it to use the selected model's default. A lower value gives that account a smaller context budget, so it compacts sooner. This can help when a smaller model gives worse answers as its prompt grows, and it does not change the context length of other messaging accounts.

Expand the boundary to read its summary, injected prompt, and compaction details. While the conversation is idle, select Regenerate to create the manual boundary again or select Remove boundary. Removing it returns the earlier transcript to normal context, so check the token counter before continuing.

Use dictation, speech, and voice calls

Select Start voice recording, speak, then select Stop recording. SoAI transcribes the recording with the selected or automatic speech-to-text model and adds the text to the composer. Review it before sending. Starting dictation ends an active voice call. Hold Start voice recording for push-to-talk, then release to transcribe and send. If the draft or selected model changes, the transcript stays in the composer for review. See Voice, dictation, and speech.

Voice call

Select Start voice call for hands-free chat. Voice Call listens until you stop speaking, sends the transcription, waits for the model, and reads the answer aloud. If you speak while the assistant is answering, SoAI stops that speech and response so your new utterance can take over. Use Pause voice call, Resume voice call, or End voice call in the call dialog. Switching conversations ends the call. If a call fails, it ends and shows the available cause, such as denied microphone access, a missing or busy microphone, or unavailable browser recording support. Correct that cause before starting another call.

Voice requirements

  • Open Chat through HTTPS or localhost so the browser can grant microphone access.
  • Allow microphone permission and use a browser with media recording support.
  • Select a conversation model before starting Voice Call.

Choose automatic or specific STT and TTS models under Chat configuration > Voice. If the composer actions are missing, turn on Voice and Voice call under Chat configuration > Appearance > Input actions. The Speak message action uses the same TTS selection.

Organize conversations

The sidebar search field matches conversation titles and message text that the browser has loaded. Select Show favorites at top to bring marked conversations to the top of the list. To rename the open conversation, select its title in the header. For a sidebar conversation, select Edit on its row to change the color, rename it, archive it, export it, or change its favorite status. Select Delete on the row to remove the conversation.

To archive, clone, or delete several conversations, expand the toolbar at the bottom of the sidebar, select Select conversations, mark the rows, then choose Archive selected, Clone selected, or Delete selected. You can select one conversation and use Clone selected when you need a single copy.

What a clone contains

A clone gets its own transcript and settings, including the assistant and tool activity already recorded. SoAI copies ready direct-file attachments. It does not copy the indexed Knowledge collection, so Knowledge attachment cards in the cloned transcript show as unavailable. To use those documents in the clone, open it and add or import them again under Add attachments > Knowledge. Later messages in the two conversations are independent. Skill selections are not copied; enable the needed skills in the clone and save. Finish or stop running and queued work before cloning or deleting a conversation.

Archive and sidebar filters

Archive hides a conversation from the active list without deleting its transcript. Expand the sidebar toolbar and select Archived conversations to search, open, unarchive, rename, recolor, favorite, or delete archived items.

Automation runs and messaging-account conversations can also appear in the sidebar. Their execution settings belong to the Automation or Messaging configuration that created them, so Chat hides or locks those controls. To remove either type from the sidebar, open Chat configuration > Appearance, turn on Hide automation runs or Hide messaging conversations, then select Save. See Connect messaging apps to SoAI to add or manage a messaging account.

A message sent from Chat while a messaging conversation is open, and the immediate reply, stay in Chat. They are not copied to the messaging app, but both remain in the transcript and can affect the model's later replies. To change that conversation's model, agent mode, workspace, display names, parameters, MCP tools, account credentials, or sender policy, select the Settings button at the right of the header, choose Settings, select Messaging, and use Edit for the account.

Files sent from Discord, Telegram, or WhatsApp appear with the provider message in this conversation. SoAI processes them as conversation attachments before the model answers. See Attachments and file context for how documents and media become model context.

You can attach files to a message that you send from Chat in a messaging conversation. That message, its files, and SoAI's immediate reply stay in SoAI. They can affect later model replies but are not forwarded to the messaging app.

To use the conversation model's default context length, open Chat configuration > Parameters, clear Context length (tokens), and select Save. A custom value must be a positive whole number.

The 64k, 96k, 128k, and 256k buttons beside the context length fill in a common value. You can still type a different value. Choose a length your model supports, then select Save; a larger context can use more memory.

Choose where persistent context belongs

Chat has several ways to keep context. Pick the narrowest scope that matches the work so unrelated conversations do not receive unnecessary instructions or data.

Context Scope Use it for
Message text or direct attachment One message in one conversation Write it in the composer or use Add attachments for a question, instruction, image, or file needed at that point in the transcript.
Conversation system prompt Model requests in the conversation Set behavior, role, format, or standing instructions under Chat configuration > Parameters > System prompt. Lock system prompt applies the same prompt to new messages and conversations until you unlock it.
Knowledge Indexed documents for one conversation Build a collection that the assistant should search across repeated questions under Add attachments > Knowledge.
Memory and profile Your user account across chats Store preferences, goals, relationships, and recurring personal context under Chat configuration > Memory. Select Add memory for a titled note or Open profile form for structured profile facts.
Workspace folder File base for one conversation Select the folder that file and tool workflows should use under Chat configuration > Files > Change folder.

Use skills for a reusable task procedure that the assistant should load when needed. Choose them under Chat configuration > Skills; they belong to the current conversation. A Skills preset can stage the same selection in another chat, but new chats and clones do not inherit it automatically.

Configure Chat

Select Chat configuration in the header, choose a tab, make your changes, then select Save. The table below separates settings that belong to the open conversation from those that apply to your WebUI or to new chats.

Tab Controls Main scope
Appearance Text zoom, widescreen layout, Markdown rendering, inline multimedia previews, automatic titles, activity display, header and composer auto-hide, sidebar visibility filters, completion and error notifications, microphone sounds, composer action visibility, and Auxiliary mobile input action, which keeps one extra action beside Send on compact layouts. WebUI preferences.
Files Select Change folder to override the workspace folder for this conversation, or Inherit default to use your default folder. SoAI removes an override if a later default folder change places it outside your allowed workspace. Current conversation.
General Primary and comparison models, Hide model id, Your name, Your photo, Model name, Model photo, Export to PDF, and Use last settings for new chats. Conversation identity plus user defaults and export preference.
Knowledge RAG enablement, embedding model, retrieval strategy, Top K, similarity threshold, and chunking controls. Use Add attachments > Knowledge for documents and reindexing. Current conversation.
Memory Search and manage saved facts, relationships, and your personal profile. See Manage memory. User memory shared across chats.
Parameters System prompt, Lock system prompt, SoAI system prompt, Context length (tokens), Reasoning effort, Max completion tokens, Agent max iterations, temperature, Top P, frequency and presence penalties, Stop sequences, Top log probabilities, and Service tier. Reasoning effort, Max completion tokens, Top P, both penalties, Stop sequences, and Top log probabilities have a Send this parameter to the model switch. Current conversation, model requests, and the user's prompt-lock preference.
Skills Import, view, download supporting files, delete, and stage the skills this chat can use. Save applies the selection. Account library and current conversation selection.
Presets Search, create, apply, rename, replace, and delete reusable configuration presets. A preset records only the sections you select, including an optional Skills selection. Parameter reset leaves model, identity, Files, Knowledge, Skills, and Tools unchanged. User preset library.
Tools Tool enablement, approval requirement, defaults for new chats, server groups, tool search, and separate Chat, Plan, and Execute selections. Current conversation plus user defaults for new chats.
Voice Automatic or specific text-to-speech and speech-to-text models. User voice preference.

After you change models, SoAI may mark a saved option such as Reasoning effort as unsupported. Choose a value supported by the new model before turning on Send this parameter to the model for that option. Automation and messaging conversations can lock settings controlled by their source configuration.

Saved prompts and configuration presets solve different problems. A saved prompt inserts reusable text into the composer. Applying a preset copies its selected sections into the Chat configuration dialog; review them and select Save. Arrow-key prompt history contains text you submitted before and remains separate from both.

Manage memory

Open Chat configuration > Memory to search your saved memories. Select Add memory to enter a title and text, or use a card's View, Edit, or Delete action. Select its observation or relationship count to manage individual records. Editing a memory with several facts also opens that record manager.

Deleting a fact preserves its memory and relationships. Deleting a memory asks you to confirm the affected facts and relationships. Use Open profile form to edit structured personal facts; deleting that profile's memory leaves separately saved user and assistant display names intact.

Select Refresh memory if entries are out of date. Review newer data before saving a conflicting edit. If a save cannot be confirmed, keep your draft and refresh to check the saved state before retrying.

Export a conversation

  1. Wait until the conversation has at least one message and its current response has finished.
  2. Open Chat configuration > General. Turn Export to PDF on for PDF or off for JSON, then select Save.
  3. Select Export conversation in the header or under More actions.

PDF export creates a styled transcript with comparison turns, tool activity, attachments, and compaction boundaries. JSON downloads the structured conversation with its model and parameter settings.

If PDF creation fails after SoAI has prepared the conversation, SoAI reports the failure and downloads the JSON version. Export stays unavailable during a response so the download does not omit a reply that is still arriving.

Leave, return, and work in another window

Switching conversations or leaving Chat does not stop work that SoAI has accepted. The sidebar marks conversations that are running or need your attention, and the timeline catches up when you return. Turn completion and error notifications on under Chat configuration > Appearance. Select Stop before leaving when you want to end the work.

When you return to Chat without choosing a conversation, SoAI reopens the last conversation available for your account in that browser and restores its saved scroll position. If that conversation is unavailable, it opens the most recent available conversation.

After a connection returns, SoAI reloads drafts, queued prompts, messages, tool activity, and attention notices. The same items appear in another signed-in browser, so a prompt queued in one browser can appear in another. You can cancel a pending queue item. Once it starts, use Stop.

Detached Chat

Select Open in new window in the top bar to detach Chat. SoAI saves the current draft and opens the selected conversation in a separate window. The detached window uses the same signed-in session and conversations.

Common blocked states

What you see What it means What to do
A draft conflict appears The same conversation draft changed in another client. Review the text and attachments SoAI kept in this browser, then continue editing or send that version.
A provider error appears with a Retry button The model provider rejected or could not complete the request. The error explains the cause and what to change, followed by the provider's own explanation when it gives one. Make the suggested change, wait for any stated delay, then select Retry to regenerate the response.
A response ends with a notice instead of an answer The model returned reasoning without a final answer, or the response reached the output token limit. Increase the max tokens setting or lower the reasoning effort, then regenerate the response. For a reasoning-only reply, retry or choose a different model or reasoning setting.
A stream error appears in the timeline The model, backend, connection, tool sequence, or system resource failed. A server error can include a reference ID. Check that the selected model is available and that its backend is running, then resend the message. Follow any action shown with the error. Before resending a prompt that used tools, inspect the tool activity: a tool may have completed even though the stream ended with an error. Repeat it only when the action is safe to run twice. If the problem continues, give the reference ID to an administrator.
Clone or delete fails because the conversation is active The conversation still has queued or running work that must reach a final state first. Resolve interactions, cancel pending queue items, and stop or wait for active work.
Context occupancy is near the model limit The transcript, prompt, attachments, and tool history are consuming most of the context window. Compact the conversation while idle, reduce input, or choose a model with a suitable context window.
Export or message history actions are unavailable A response or agent turn is still running. Wait for it to finish or select Stop, then try the action again.
Queued prompts show Paused A response was stopped while these prompts were waiting, so SoAI held them instead of starting them. Leave the composer empty and select Resume queued prompts, send another message to run it first, or cancel the items you do not need.
Send is unavailable You have not selected an available model, the composer is empty, or an attachment is not ready. Select an available model, add input, and wait for processing or remove the failed attachment.
Tool or mode controls are locked The conversation is running, Plan or Execute requires tools, or another feature manages required tools. Finish or stop the turn. To turn tools off, switch to Chat mode and disable indexed Knowledge first if it is keeping the required Knowledge tools enabled.
Voice controls are disabled The page is not in a secure context, the browser lacks media support, permission was denied, or no model is selected. Use HTTPS or localhost, allow microphone access, check the input device, and select a model.

Related documentation