Frequently asked questions (FAQ)
Common questions in one place: on-prem and air-gapped operation, where data is stored, supported formats, roles and permissions, embedding agents, and language support.
This page groups the questions that come up most often when evaluating or running Flux. It covers how it is deployed, where your data is stored, which files you can upload, how role-based permissions work, how to embed an agent on an external site, and which languages are supported. For step-by-step procedures, see the dedicated guide for each topic as well.
Deployment and operations
Flux is an on-premises platform that you install and run on your own servers. The default setup runs the application together with its databases, vector and graph stores, and object storage on a single server using Docker Compose.
You install and run it yourself. With the provided Docker Compose setup, Flux and all of its supporting infrastructure (relational database, vector DB, graph DB, object storage, cache, and message queue) run together on one server inside your own environment.
Yes. Flux is designed to handle core work such as embedding, search, and document parsing with self-hosted models, so a closed-network configuration can operate without an external internet connection. Note that if you choose an external LLM, embedding, or speech-to-text provider, communication with that provider's API is required.
No. Text and image embeddings and speech-to-text use self-hosted models built into the server by default. The answering LLM is connected by registering a provider in organization settings, where you can connect a model hosted on your own infrastructure or pick an external provider.
Whether you can run fully air-gapped depends on which components you self-host. If the answering LLM is also an in-house model, you can operate without external traffic; using any external provider introduces communication with that provider's API.
Data storage and security
Uploaded documents, conversations, and settings are all kept inside the company infrastructure where Flux is installed. Data is split across the stores below by type.
Relational database
Structured data such as organizations, users, document metadata, conversation history, and settings is kept in PostgreSQL.
Object storage
Original uploaded files and extracted images are kept in S3-compatible object storage (MinIO).
Vector search store
Document embeddings (vectors) used for search are kept in a vector database (Milvus), and every search is isolated by organization so no other organization's vectors are mixed in.
Knowledge graph store
Entities and relationships extracted from documents are stored in a graph database (Neo4j) and used for knowledge graph search.
In a setup that uses only built-in models, document data never leaves your infrastructure. If you connect an external LLM, embedding, or speech-to-text provider, the data needed for that feature is sent to the provider's API when the feature runs. You choose which external providers, if any, to use in organization settings.
All data is isolated per organization (tenant). Documents, conversations, agents, and search results are always filtered to the requesting user's organization, and other organizations' data cannot be reached.
External provider API keys registered for an organization are stored encrypted. The full key is not shown again in plain text in the settings screen.
Organization isolation is Flux's primary security boundary. Every request passes authentication and organization-membership checks and only reaches data in the verified organization.
Supported file formats
Flux lets you upload not only documents but also images and audio and make them searchable. Each type is handled differently.
| Category | Formats | How it is handled |
|---|---|---|
Documents | PDF, Word (DOC, DOCX), PowerPoint (PPTX), Excel (XLSX) | Body text and table structure are extracted and indexed as text. |
Text | TXT, Markdown (MD), HTML | The content is read as-is and indexed. |
Images | PNG, JPG, WebP, GIF, TIFF | Text is extracted with OCR, and a generated caption describes the image so it can be searched. |
Audio | MP3, WAV | Speech-to-text transcribes what was said into text and indexes it. |
The upload size limit is 1GB per file. OCR, captioning, and speech-to-text can be turned on or off in organization settings.
Roles and permissions
What you can do inside an organization depends on your role. Roles run from least to most privileged as Viewer, Member, Admin, and Owner, and higher roles include every permission of the lower ones.
| Permission | Viewer | Member | Admin | Owner |
|---|---|---|---|---|
| Search and view | ||||
| Add and edit documents | ||||
| Delete documents | ||||
| Add and edit agents | ||||
| Invite and manage members | ||||
| Change organization, AI, and RAG settings | ||||
| Delete the organization |
One person can belong to several organizations and may have a different role in each. Use the organization switcher at the top left to move between them.
Agents and embedding
An agent built in the Agent Builder can be used in chat inside your company, and it can also be shared externally as a public link. Embedded chat is usable by anyone without logging in, so turn it on only for agents you are comfortable making public.
Publish the agent
In the builder, set the agent to published. Only published agents can be selected for chat and embedding.
Enable embedding
In the agent's embed settings, turn on embedding and set the identifier (slug) used in the external address along with an intro description.
Share the public link
Open the issued embed address in a new tab or link to it from an external page or internal portal, and visitors can start chatting right away.
No. Embedded chat works without authentication. It can reach only agents that have both embedding and publishing turned on, and a request-rate limit is applied to prevent abuse.
Yes. Embedded conversations are saved as records and tracked as anonymous users. Embedded conversations that go unused for a period are cleaned up automatically.
Turning on embedding makes a published agent available externally. For agents that handle sensitive internal information, keep embedding off, or connect only material that is safe to share publicly.
Language and search quality
The interface language supports Korean and English and can be switched from the menu at the top right. Document search uses indexing tuned for both Korean and English content.
The bracketed numbers in an answer are citations that point to the source a statement is based on. Clicking a number jumps to the matching source card so you can check the original location.
First check whether the document has finished indexing. If results are still lacking, try making your question more specific, or verify that the agent's retrieval tool is set to look at the correct document group.
Yes. For images, text is extracted with OCR and a generated caption describes the picture, and both are indexed. As a result, wording or scenes inside a photo can also turn up in search.
When you are stuck
If you did not find what you were looking for, use search within the help center to quickly locate another guide. For operational questions, contact your organization administrator or the team that operates Flux.
- Open search in help (macOS)⌘K
- Open search in help (Windows)CtrlK