Modules
Chat
Conversations with models, agents, and flows, with attachments, knowledge search, questions about tables, and generated files.
On this page
Chat is where people in your organization talk to models, agents, and published flows, about their own files and the workspace's knowledge and data contexts.
Start a conversation
- Open Chat and start a new conversation.
- Choose the target: a model enabled in the workspace, a published agent, or a flow published to chat.
- Write the message or attach files. The answer streams in.
The target is locked after the first message. Only the models, agents, and flows available to that workspace and that person appear.
Attachments
Chat accepts four kinds of attachment, up to 100 MB per file: documents, images, audio, and video. The administrator sets which kinds are enabled and the maximum size of each.
| Kind | How it is read |
|---|---|
| Documents | Text extracted from txt, md, csv, tsv, json, xml, html, log, digital PDF, docx, xlsx, and pptx. The legacy .doc, .xls, and .ppt formats are accepted, but their text is not extracted. |
| Images | Sent directly to vision-capable models. For other models, a vision model chosen by the administrator describes the image. |
| Audio | Transcribed by a transcription model. |
| Video | The audio track is transcribed; the visual content is not analyzed. |
Attached files go into the sender's personal area, indexed and visible only to that person, and can be used in later conversations.
Voice input
Voice input uses the browser's speech recognition or a transcription model on the server, as configured by the administrator.
Web search
Web search is switched on per message. The default provider is DuckDuckGo; Tavily and Brave work with your own key. The query is sent to the chosen search provider. If search is unavailable, the answer says so.
Answers from knowledge
With a knowledge context selected, the answer is grounded by hybrid search: vector search in Milvus and text search in PostgreSQL, merged with Reciprocal Rank Fusion and reordered by a rerank model. The answer shows the sources it used.
Questions about tables
With a data context selected, the model writes a single SELECT query, validated and run in read-only mode, with up to 40 rows per answer. The answer names the context it used and lets you see the SQL behind it.
Ad-hoc SQL is off by default in every data context. For chat to answer from a context, the person who manages it turns on Allow ad-hoc SQL in Storage; the same option enables manual queries. While it is off, neither manual queries nor chat can query the context.
Generated files
From a conversation, chat generates files for download:
| Format | Content |
|---|---|
| .xlsx | A native spreadsheet, with the data in cells. |
| .pptx | A presentation built from the Markdown headings. |
| A PDF document. | |
| .csv, .md, and .txt | Tabular data and text. |
The .docx format is not offered.
Long conversations
In long conversations, the platform recalls relevant earlier turns by similarity. When part of the history has to be left out to fit the model's context window, the model is told so.
Controls and provenance
- Stop the generation in progress.
- Rate the answer up or down and copy it.
- Open the inspector, which shows the exact request sent to the model and the answer's provenance.
- Rename and delete conversations. Deletion is recorded in the audit trail.
Each answer keeps the model that produced it and the sources that grounded it: documents with scores, or the SQL with the number of rows. See Governance and compliance.